Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries

# Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries
Source: [https://arxiv.org/html/2607.10113](https://arxiv.org/html/2607.10113)
###### Abstract

Large language model agents increasingly store reusable procedures outside the model\. These reusable procedures are often called*skills*: they may be code functions, natural\-language instructions, SKILL\.md packages, workflow graphs, or learned adapters that a future agent can retrieve and invoke\. This taxonomy\-driven survey asks how such skill libraries change over time\. Across a 124\-paper 2023–2026 audit set, we synthesize dynamic skill systems as*lifecycle\-managed, verified, evolving artifact stores*: agents collect evidence from interaction, propose skill updates, verify and admit candidates, organize them for retrieval and composition, repair or prune stale entries, and govern sharing through provenance and rollback\. We organize the literature around three survey tools\. First, a six\-sense taxonomy distinguishes the structurally different artifacts called “skills” in current papers\. Second, an eight\-stage lifecycle architecture identifies the recurring design decisions behind evidence acquisition, proposal, verification/admission, storage, retrieval/composition, maintenance, distillation/portability, and governance\. Third, a lightweight skill\-record schema and ten\-operator vocabulary provide common terms for comparing library updates without elevating them into a separate method contribution\. Using this structure, we synthesize evidence\-graded patterns with explicit caveats: admission and repair are repeatedly important, verifier quality materially affects skill\-aware RL, flat retrieval can degrade as libraries grow, and current benchmarks still under\-report library trajectories, usage–utility gaps, and safety surfaces\. We close with concrete reporting standards and open problems for evaluating dynamic skills as changing libraries rather than static prompt or tool collections\.

## 1Introduction

Large language model agents are moving from chat\-style assistance into operational workflows\. They browse and manipulate web interfaces, write and repair software, operate desktop and computer\-use environments, coordinate multi\-step research pipelines, interact with embodied environments, and assist domain\-specific workflows such as recommendation and medical imaging\. Across these settings, the bottleneck is not only whether the model can reason about a single task\. It is whether the agent can reuse procedural knowledge across recurring tasks, tools, interfaces, and users instead of re\-deriving the same strategy inside every context window\.

*Agent skills*were proposed as a practical answer to this reuse problem\. In plain terms, a skill is a reusable procedure that an agent can call later\. It may be a function, a natural\-language instruction, a SKILL\.md directory, a workflow graph, or a learned adapter, but its role is the same: preserve a reusable way of acting so that a future agent can retrieve, compose, and execute it\. Early executable libraries such asVoyagerandLATMstored code skills; 2025 systems extended skills into web and software agents; 2026 systems expanded the ecosystem into SKILL\.md packages, computer\-use skills, mobile GUI skills, multimodal skills, benchmarks, registries, and safety audits\(Wanget al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib1); Caiet al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib2); Zhenget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib13); Xiaet al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib19); Chenet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib28);[b](https://arxiv.org/html/2607.10113#bib.bib103); Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12); Xieet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib113); Taoet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib97); Fanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib99)\)\. The broader skill ecosystem now includes SKILL\.md conventions, function\-calling schemas, MCP/plugin\-style packaging, and registry\-like platforms that support libraries with hundreds or thousands of reusable artifacts\(Jianget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib54); Zhanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib115); Anthropic,[2025](https://arxiv.org/html/2607.10113#bib.bib116); Liet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib65); Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43)\)\.

The word “skill” is overloaded\. Table[1](https://arxiv.org/html/2607.10113#S1.T1)is therefore the conceptual entry point for the survey: it distinguishes executable programs, natural\-language lessons, SKILL\.md packages, parametric adapters, memory traces, and capability labels before the paper introduces lifecycle or update notation\. The first four senses form the main dynamic\-skill cluster studied in this survey; the latter two are boundary cases included when they illuminate library behavior\.

MethodSense of “skill”Artifact formExec\.EditPortableInspect\.Verif\. handleVoyager\(Wanget al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib1)\)Executable code\.pylibrary✓✓∼a\\sim^\{\\mathrm\{a\}\}✓unit testLATM\(Caiet al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib2)\)Executable codePython tool✓✓∼a\\sim^\{\\mathrm\{a\}\}✓unit testAgentFactory\(Zhanget al\.,[2026i](https://arxiv.org/html/2607.10113#bib.bib14)\)Executable codePython \+ MCP✓✓∼a\\sim^\{\\mathrm\{a\}\}✓meta\-agent inspect\.LIVE\-SWE\(Xiaet al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib19)\)Executable codecode snippet✓✓∼a\\sim^\{\\mathrm\{a\}\}✓rolloutSAGE\(Wanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib3)\)Executable codePython function✓✓∼a\\sim^\{\\mathrm\{a\}\}✓env\. rewardASI\(Wanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib100)\)Executable codeprogram skill✓✓∼a\\sim^\{\\mathrm\{a\}\}✓rolloutHASP\(Liuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib124)\)Executable codeprogram function✓✓∼a\\sim^\{\\mathrm\{a\}\}✓teacher/rolloutSkillOps\(Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\)Executable contracttyped skill contract✓✓∼a\\sim^\{\\mathrm\{a\}\}✓validator/graph auditERL\(Allardet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib7)\)NL heuristiccritique–✓✓✓indirectRetroAgent\(Zhanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib8)\)NL heuristicdual lesson–✓✓✓indirectEvolveR\(Wuet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib11)\)NL heuristicstrategic principle–✓✓✓indirectMetaClaw\(Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35)\)SKILL\.mdL1/L2/L3 markdown✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓L3 code gateMemento\(Zhouet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib10)\)SKILL\.md \+ casemarkdown \+ trace✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓L3 code gateTrace2Skill\(Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22)\)SKILL\.mdL2 markdown✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓rubric judgeAutoRefine\(Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25)\)SKILL\.mddual\-form markdown✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓rubric judgeSkillFlow\-Bench\(Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48)\)SKILL\.md patchfiles \+ JSON patch✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓benchmark verifierSkillOpt\(Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128)\)SKILL\.md / text skilloptimized document–✓✓✓held\-out utilitySkillGrad\(Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130)\)SKILL\.mdstructured package✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓loss/validationMUSE\-Autoskill\(Linet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib129)\)SKILL\.md \+ memoryskill \+ per\-skill memory✓b\\checkmark^\{\\mathrm\{b\}\}✓✓✓unit testsContractSkill\(Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102)\)Contract skillpre/postcondition artifact✓✓✓✓deterministic checksABSTRAL\(Songet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib29)\)SKILL\.md / design docstructured markdown–✓✓✓trace evidenceSkillsCrafter\(Wanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib33)\)ParametricLoRA subspace✓c\\checkmark^\{\\mathrm\{c\}\}–––behavioral probeSKILL0\(Luet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib31)\)Parametricadapter✓c\\checkmark^\{\\mathrm\{c\}\}–––behavioral probeSELAUR\(Zhanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib30)\)ParametricLoRA✓c\\checkmark^\{\\mathrm\{c\}\}–––behavioral probeLSE\(Chenet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib32)\)Parametric \(mix\)prompt \+ weight✓c\\checkmark^\{\\mathrm\{c\}\}∼\\sim–∼\\simbehavioral probeSimpleMem\(Liuet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib58)\)Memory / trajectorytrace case–∼\\sim✓✓indirectMUSE\(Yanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib9)\)Memory / trajectoryepisodic store–∼\\sim✓✓indirectCASCADE\(Huanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib20)\)Memory / trajectorycurated case set–∼\\sim✓✓indirectXSkill\(Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12)\)Skill \+ experienceMD skill \+ JSON exp\.∼\\sim✓✓✓indirectSkillFlow\-2025\(Liet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib47)\)Registry\-retrieved skillSKILL\.md corpus \+ ranked candidates∼\\sim–✓✓retrieval evalSRA\(Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)Registry\-retrieved skilllarge skill corpus∼\\sim–✓✓retrieval \+ utility evalSkillsVote\(Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125)\)Registry\-governed skillexecutable \+ guidance corpus∼\\sim✓✓✓evidence gateSSL\(Lianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib112)\)Structured representationJSON\-like skill graph–✓✓✓support auditCo\-Evolving\(Junget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib36)\)Capability labeldeclarative role–∼\\sim✓✓noneGEA\(Wenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib51)\)Capability labeldeclarative role–∼\\sim✓✓none

Table 1:Representative dynamic\-skill methods keyed by paper, with the sense of “skill” each adopts and the five properties that determine its dynamic behaviour\.The six senses partition the design space \(horizontal rules\) and project directly onto the taxonomy axes of Section[6](https://arxiv.org/html/2607.10113#S6); the first four senses are the “dynamic skills” studied in this survey while the last two are boundary cases with restricted edit / verification dynamics\. Column definitions:*Exec\.*= is the artifact machine\-executable;*Edit*= can be revised without retraining;*Portable*= usable across different LLM backbones;*Inspect\.*= auditable in textual form;*Verif\. handle*= form of admission check available\. Symbols:✓\\checkmark= fully satisfies;∼\\sim= partial / indirect; – = does not satisfy\.aportable across backbones that can call the same runtime;bexecutable only when L3 attaches code or MCP resources;cexecutable only through the compatible base model\.This adoption changes the research question\. In OpenClaw\-style personal agents,111We use OpenClaw,SkillClaw, andClawSafetyas research\-artifact or evaluation\-setting names from the cited papers; the survey relies on the papers’ methods and measurements, not on product or platform claims\.skills are no longer helpful prompt snippets; they are part of the execution substrate\.SkillClawuses cross\-user OpenClaw interactions to evolve shared skills over deployment rounds\(Maet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib23)\), whileClawSafetyshows that skill files can become a high\-trust prompt\-injection channel in privileged personal\-agent scaffolds\(Weiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib56)\)\. The same abstraction that improves reuse creates questions of verification, maintenance, provenance, and governance\.

Many skill libraries are still treated as*static*: authored once, versioned rarely, and weakly connected to the trajectories that reveal whether a skill remains useful, safe, or correct\. Recent work\(Alzubiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib15); Zhenget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib13); Wanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib100);[2026g](https://arxiv.org/html/2607.10113#bib.bib101); Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102); Chenet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib103); Wanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib3); Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22); Shiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib17); Liet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib42); Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132); Liet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib47); Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35); Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128); Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130); Linet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib129)\)rejects this assumption by treating the library itself as a learning object that grows, shrinks, repairs itself, changes storage structure, and sometimes distills its contents back into model weights\. The methods differ—retrospective consolidation, execution\-grounded honing, RL\-based proposal, rollback\-validated refactoring, registry\-scale retrieval, structured repair, library\-time technical\-debt maintenance, text\-space skill optimization, and two\-timescale distillation—but share the commitment that skill libraries should be managed as evolving systems\.

This survey synthesizes the field through a lifecycle view: dynamic skill systems are*lifecycle\-managed, verified, evolving artifact stores for LLM agents*\. A skill artifact is created from evidence, proposed as an edit, verified, admitted, organized, retrieved or composed, maintained, sometimes distilled, and governed through provenance and rollback\. Papers differ mainly in which stages they implement, which learning signal drives edits, and where they place the verifier\.

The result is a taxonomy\-driven survey rather than an encyclopedic paper\-by\-paper catalog: the goal is to make artifact boundaries, lifecycle commitments, update mechanisms, and evidence strength comparable across a fast\-moving literature\.

### 1\.1Dynamic skill libraries as learning systems

A*library*is a collection of such skills equipped with a retrieval or composition mechanism\. A*dynamic*library is one whose contents or organization can change as the agent accumulates evidence\. We use a compact record schema and operator vocabulary later in the paper to make this comparison precise, but the basic idea is simple: a dynamic skill system does not only choose which skill to read at inference time; it changes the store that future agents will read from\.

This makes dynamic skills a learning\-systems problem rather than merely a software\-packaging problem\. Interaction produces evidence; evidence selects edits; edits change the library; the changed library alters the distribution of future actions through retrieval, composition, and execution\. In this view,ℒt\\mathcal\{L\}\_\{t\}is part of the agent’s state, the transitionℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}is a learning rule, verification is a selection mechanism, maintenance is a form of regularization, and distillation is a consolidation step that moves external procedural knowledge into slower parametric or semi\-parametric stores\. The same library can also serve as the interface between individual learning and collective learning when skills are shared across users, registries, or agents\.

The time index matters because deployed libraries go stale as tasks, tools, backbones, and safety constraints drift\. It also makes the evidence\-graded patterns in §[9](https://arxiv.org/html/2607.10113#S9)expressible: admission and verifier quality, retrieval degradation, weaker\-backbone gains, focused curation, maintenance at scale, and write\-time abstraction are all properties of a changing library\.

### 1\.2Why a survey now

Three developments in 2025–2026 made dynamic skills surveyable\. First, packaging and registry infrastructure matured enough that libraries must be curated, routed, and governed as collections \(§[10\.3](https://arxiv.org/html/2607.10113#S10.SS3)\)\. Second, methods now span the learning\-signal spectrum, from LLM\-as\-judge admission through audited verification and two\-timescale RL\. Third, deployment and benchmark papers report failure modes—retrieval degradation, skill injection, verification drift, skill inflation, and maintenance\-off collapse—that earlier method papers treated mostly in ablation\.SkillFlow\-Bench\(Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48)\), for example, evaluates discovery, patching, reuse, and maintenance over sequential tasks and shows that high skill\-use rates need not imply high utility\.

The field still lacks shared vocabulary and benchmark protocol\. Papers use “skill” for at least six structurally different artifacts; update operations are named inconsistently; benchmarks often report terminal performance rather than library trajectories \(§[8](https://arxiv.org/html/2607.10113#S8)\)\. The survey’s role is to provide a common language before ranking is meaningful\.

This paper is complementary to prior surveys\.Xu and Yan \([2026](https://arxiv.org/html/2607.10113#bib.bib67)\)provide a broad account of agent skills across architecture, acquisition, security, and future directions;Fanget al\.\([2025](https://arxiv.org/html/2607.10113#bib.bib68)\)survey self\-evolving agents beyond skill libraries;Zhenget al\.\([2025c](https://arxiv.org/html/2607.10113#bib.bib69)\)emphasize lifelong LLM\-agent learning with a memory\-centric lens; andZhouet al\.\([2026b](https://arxiv.org/html/2607.10113#bib.bib114)\)provide a broad May 2026 taxonomy of agent\-skill techniques and applications\. Our focus is narrower: dynamic, lifecycle\-managed skill libraries and the mechanisms by which such libraries change over time\.

### 1\.3Contributions

This paper makes five contributions\.*First*, it clarifies what current papers mean by “skill” through the six\-sense taxonomy in Table[1](https://arxiv.org/html/2607.10113#S1.T1)\.*Second*, it recasts dynamic skills as editable learning\-system artifacts and introduces a lightweight skill\-record schema plus library\-level transition notation for comparing systems \(§[3](https://arxiv.org/html/2607.10113#S3)\)\.*Third*, it introduces a lifecycle architecture spanning evidence acquisition, proposal, verification/admission, organization, retrieval/composition, maintenance/repair, distillation/portability, and governance \(§[5](https://arxiv.org/html/2607.10113#S5)\)\.*Fourth*, it uses a ten\-operator vocabulary \{Add,Refine,Merge,Split,Prune,Distill,Abstract,Compose,Rewrite,Rerank\} to organize mechanisms and system families without claiming closure or composition laws \(§§[6](https://arxiv.org/html/2607.10113#S6)–[7](https://arxiv.org/html/2607.10113#S7)\)\.*Fifth*, it audits lifecycle\-aware evaluation, then distills seven evidence\-graded patterns, eight safety surfaces, and eight open problems tied to concrete next experiments \(§§[8](https://arxiv.org/html/2607.10113#S8),[9](https://arxiv.org/html/2607.10113#S9),[11](https://arxiv.org/html/2607.10113#S11),[12](https://arxiv.org/html/2607.10113#S12)\)\.

## 2Corpus, Scope, and Review Protocol

We separate the object of study from the process used to assemble it\. The survey covers papers that treat skill*libraries*or skill\-like external artifacts as dynamic objects: artifacts may be added, revised, merged, pruned, routed, distilled, transferred, or governed as the agent accumulates evidence\. It excludes classical option discovery, generic tool\-use benchmarking, and fine\-tuning pipelines unless they produce an externally invocable skill artifact or directly evaluate such artifacts’ lifecycle\.

### 2\.1Inclusion Criteria

The working audit set contains 124 modern papers after the May 31, 2026 update: 2 from 2023, 19 from 2025, and 103 from 2026\. This set includes the primary dynamic\-skill method, benchmark, infrastructure, and safety cluster, plus boundary/context papers that define terminology, adjacent lifelong\-agent evaluation, or neighboring self\-evolution settings\. Older classical\-options and hierarchical\-RL anchors are cited for comparison but are not counted in this modern audit set\. We included papers from 2023–2026 when they satisfied at least one of three criteria: \(i\) the paper proposes a mechanism that adds, edits, prunes, distills, routes, transfers, or composes an agent skill artifact; \(ii\) the paper provides infrastructure or a benchmark that changes how skill libraries are stored, retrieved, evaluated, or governed; or \(iii\) the paper documents a safety or deployment failure mode specific to skill artifacts\. Boundary/context works — the systematization\-of\-knowledge \(SoK\) paper on agentic skills\(Jianget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib54)\),Xu and Yan \([2026](https://arxiv.org/html/2607.10113#bib.bib67)\)’s 2026 position paper,MAGELLAN’s autotelic curriculum mechanism\(Gavenet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib64)\), theELL/StuLifebenchmark\(Caiet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib71)\), and theExperience Compression Spectrumframework\(Zhanget al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib94)\)— are retained as scope\-setting evidence rather than as members of the primary method cluster\.

Figure[2](https://arxiv.org/html/2607.10113#S6.F2)visualizes the corpus at the level needed for survey synthesis: a sparse 2023 foundation, no primary 2024 item after filtering, a 2025 transition wave, and a dense 2026 expansion into executable libraries, infrastructure, benchmarks, and safety\.

### 2\.2Search and Screening Procedure

The corpus was assembled by iterative database search and backward/forward snowballing rather than by a single PRISMA\-style query\. We searched arXiv, Semantic Scholar, Google Scholar, OpenReview, and venue proceedings using combinations of*LLM agent skill*,*agentic skill*,*skill library*,*self\-evolving agent*,*lifelong agent*,*skill routing*,*SKILL\.md*,*agent skill safety*,*skill benchmark*, and named\-system queries for papers discovered during snowballing\. Candidate papers were screened first by title/abstract and then by PDF when the abstract suggested a persistent artifact, skill\-like package, evolving library, or skill\-specific safety/evaluation surface\. We excluded papers whose only adaptation object was model weights, prompt optimization, generic tool use, or episodic memory without an invocable interface, unless the paper directly evaluated how such artifacts become reusable skills\. Because the collection was assembled in an arXiv\-heavy and still\-moving area, we report the final audit set and inclusion frontier rather than a PRISMA exclusion flow\. The survey does not claim an exhaustive count of every tool\-use, memory, or HRL paper adjacent to the topic\.

### 2\.3Temporal Cutoff and Update Policy

All statements in the main text should be read against an explicit temporal cutoff: May 31, 2026\. Papers appearing after that date are outside the surveyed corpus\. The cutoff is an audit boundary, not a signal that the field has stabilized\. We distinguish*corpus updates*, which add papers satisfying the criteria above, from*synthesis updates*, which revise the taxonomy, evidence\-graded patterns, or open problems only when new evidence changes a lifecycle stage or exposes a new failure mode\. Product and specification sources such as the public Agent Skills format are cited only for packaging facts and are not counted as papers in the audit set\.

### 2\.4Note Schema and Conflict Resolution

Each included paper was read into a common note schema covering problem framing, mechanism, control action, experimental setup, headline results, ablations, limitations, closest peers, and survey takeaway\. When duplicate notes disagreed, we resolved the conflict against the PDF\. The schema is important because several names are close but technically distinct:SkillFlow\-2025is a multi\-stage retrieval system for community skill repositories, whileSkillFlow\-Benchis a lifelong skill\-discovery benchmark;SAGEstores executable Python skills, whilePSN’s refactors are rollback\-validated graph edits\.

### 2\.5Evidence Roles

The corpus mixes method papers, infrastructure papers, benchmarks, and safety studies\. We use method papers for evidence about operators, triggers, verifiers, and mechanisms; benchmark papers for evaluation protocol and lifecycle failure modes; infrastructure papers for storage, portability, registries, and operator cost; and safety papers for attack surfaces and partial defenses\. We grade evidence qualitatively:Ameans multiple controlled ablations or one clean ablation plus independent corroboration;Bmeans one controlled study or a strong benchmark/deployment measurement;Cmeans convergent benchmark behavior without a clean causal ablation; andDmeans architectural corroboration only\. We avoid cross\-paper leaderboards unless the harness is shared, because backbone, tool surface, context budget, and evaluator often change together\.

### 2\.6Relation to Adjacent Surveys

Four peer surveys cover adjacent territory and we do not duplicate their full scope\.Xu and Yan \([2026](https://arxiv.org/html/2607.10113#bib.bib67)\)surveys agent\-skill architecture, acquisition, and security from a broad position\-paper vantage;Fanget al\.\([2025](https://arxiv.org/html/2607.10113#bib.bib68)\)surveys self\-evolving agents beyond skill libraries;Zhenget al\.\([2025c](https://arxiv.org/html/2607.10113#bib.bib69)\)surveys lifelong learning for LLM agents with a memory\-centric emphasis; andZhouet al\.\([2026b](https://arxiv.org/html/2607.10113#bib.bib114)\)provides a broad May 2026 taxonomy of agent\-skill techniques and applications\. Our contribution relative to these is the library\-as\-object framing, the lifecycle architecture, and the operator\-level vocabulary: instead of asking whether agents can learn over time in general, we ask how an externally invocable artifact store changes, verifies, maintains, and governs itself\.

The scope also separates this survey from classical options and hierarchical reinforcement learning\. Options, skill chaining, option\-critic methods, and deep skill\-discovery methods study temporally extended policies inside an environment\(Suttonet al\.,[1999](https://arxiv.org/html/2607.10113#bib.bib59); Konidaris and Barto,[2009](https://arxiv.org/html/2607.10113#bib.bib60); Baconet al\.,[2017](https://arxiv.org/html/2607.10113#bib.bib61); Eysenbachet al\.,[2019](https://arxiv.org/html/2607.10113#bib.bib62); Nachumet al\.,[2018](https://arxiv.org/html/2607.10113#bib.bib63)\)\. Dynamic agent skills add a different object: an externally inspectable artifact that can be edited after deployment, packaged with metadata, admitted or rejected by language\- or execution\-level verifiers, shared across agents, and governed through provenance\. We use the options framework as a starting point because it clarifies applicability, policy, and termination; the survey’s added components are the artifact\-store machinery that classical HRL usually leaves implicit\.

## 3Skill Records and Library Updates

The preceding section and Table[1](https://arxiv.org/html/2607.10113#S1.T1)establish the terminology\. This section gives the lightweight notation used in the rest of the survey\. The notation is a comparison scaffold, not a separate model of skill learning: it lets us say which artifact is being edited, which verifier admits it, which lineage record survives, and how the library changes from one state to the next\.

### 3\.1The six senses of “skill”

Table[1](https://arxiv.org/html/2607.10113#S1.T1)partitions the literature along five dynamic properties: whether the artifact is*executable*,*editable in place*,*portable across models*,*inspectable by humans*, and attached to a*verification handle*\. These properties largely determine which triggers, operators, and signals a method can support\.

The executable\-code sense, occupied byVoyager,LATM,LIVE\-SWE,AgentFactory,SAGE,ASI,HASP, andSkillOps\(Wanget al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib1); Caiet al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib2); Xiaet al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib19); Zhanget al\.,[2026i](https://arxiv.org/html/2607.10113#bib.bib14); Wanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib3);[b](https://arxiv.org/html/2607.10113#bib.bib100); Liuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib124); Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\), treats a skill as a named program or typed contract that can be invoked, inspected, edited line\-by\-line, and verified by tests, validators, or environment feedback\. The NL\-heuristic sense, occupied byERL,RetroAgent,EvolveR, andEmbodiSkill\(Allardet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib7); Zhanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib8); Wuet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib11); Juet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib118)\), treats a skill as a short lesson or procedural specification injected at retrieval time; it is editable and portable but only indirectly verifiable through downstream task success\. The SKILL\.md sense, occupied byMetaClaw,ABSTRAL,Trace2Skill,Memento,K2\-Agent,AutoRefine,SkillFlow\-Bench,SkillOpt,SkillGrad, andMUSE\-Autoskill\(Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35); Songet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib29); Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22); Zhouet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib10); Wuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib34); Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25); Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48); Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128); Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130); Linet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib129)\), unifies code and prose behind a progressive\-disclosure package: L1 metadata for routing, L2 instructions for the agent, and L3 code, schemas, or sub\-skills for execution and verification\. Product/specification sources describe the public packaging convention as a directory with aSKILL\.mddescriptor, optional scripts, and progressive disclosure from metadata to full instructions and resources\(Zhanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib115); Anthropic,[2025](https://arxiv.org/html/2607.10113#bib.bib116)\); we use those sources only for format facts, not as evidence for lifecycle effectiveness\.

The parametric sense, represented bySkillsCrafter,SELAUR,SKILL0, andLSE\(Wanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib33); Zhanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib30); Luet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib31); Chenet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib32)\), treats a skill as a weight delta, trained prompt module, or internalized procedure; such artifacts are probed behaviorally, ported only across compatible base models, and edited through training\. The memory/trajectory sense, present inSimpleMem,MUSE,CASCADE,XSkill,SkillTTA, and the case layer ofMemento\(Liuet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib58); Yanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib9); Huanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib20); Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12); Wanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib123); Zhouet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib10)\), blurs skills with retrievable episodes; the boundary is whether the trace has an invocable interface\. The registry\-retrieved and capability\-label senses, present inSkillFlow\-2025,SkillsVote,SkillsInjector,Co\-Evolving\-Agents, and some ofGEA\(Liet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib47); Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125); Liet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib131); Junget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib36); Wenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib51)\), are the most brittle under dynamic evolution because retrieval success or claimed capability has no robust edit body unless paired with executable skill contents and admission checks\.

In the rest of the paper, “skill” means executable code, NL heuristic, SKILL\.md package, or parametric skill by default; memory/trajectory and capability\-label artifacts are treated as boundary cases\.

This definition also separates*skills*from ordinary*tools*\. A tool is usually an externally supplied callable resource: the agent decides when to invoke it, but the tool’s body, interface, and release process are not learned by the agent\. A dynamic skill is a persistent artifact whose lifecycle is itself part of learning\. The same Python function or MCP endpoint can therefore be a tool in a generic tool\-use benchmark and a skill in a dynamic\-skill system if the agent proposes it, revises it from trajectories, verifies it for admission, stores lineage, and later maintains or transfers it\. The distinction is operational rather than syntactic: what matters is whether the artifact participates inℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}\.

### 3\.2From static skill records to editable records

#### Starting point\.

The canonical reference for skills as temporally extended actions is the options framework ofSuttonet al\.\([1999](https://arxiv.org/html/2607.10113#bib.bib59)\): an option is a triple⟨I,π,β⟩\\langle I,\\pi,\\beta\\rangleof an initiation setI⊆𝒮I\\subseteq\\mathcal\{S\}, a policyπ\\pi, and a termination conditionβ\\beta\. Within the LLM\-agent literature, the systematization\-of\-knowledge paper byJianget al\.\([2026b](https://arxiv.org/html/2607.10113#bib.bib54)\)reinterprets this triple for language agents as a four\-tuple

𝒮=⟨C,π,T,R⟩,\\mathcal\{S\}\\;=\\;\\langle C,\\;\\pi,\\;T,\\;R\\rangle,\(1\)whereCCis an applicability predicate \(“when is this skill relevant”\),π\\piis the executable policy \(code, prompt template, adapter, or lesson text\),TTis a termination condition \(success, exception, or budget exhaustion\), andRRis a*reusable interface*: the invocation name, expected inputs, outputs, preconditions, return values, and composition points that let another agent, tool, or skill call the artifact\.

#### Five concrete gaps\.

TheSoK\-Skillsfour\-tuple is adequate for*static*libraries, but dynamic libraries require five missing objects\. First, a*time index*distinguishes𝒮t\\mathcal\{S\}\_\{t\}from later refinements, replacements, and merges\. Second, an*edit operator*records whether a method rewrites NL lessons \(ERL,EvolveR,RetroAgent\), function bodies \(SAGE\), SKILL\.md prose \(AutoRefine,Memento\), symbolic refactors \(PSN\)\(Shiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib17)\), or mutation heuristics \(CODE\-SHARP\)\(Bornemannet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib6)\)\. Third, an*admission gate*captures filters such asEvoSkill’s Pareto front\(Alzubiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib15)\),SkillCraft’s MCP verifier\(Chenet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib18)\), andASG\-SI’s audited graph\(Huang and Huang,[2025](https://arxiv.org/html/2607.10113#bib.bib55)\)\. Fourth,*lineage*supports rollback, supersession, and maturity gating in systems such asAgentDevel,PSN,Memento, andTrace2Skill\(Zhang,[2026](https://arxiv.org/html/2607.10113#bib.bib50); Shiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib17); Zhouet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib10); Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22)\)\. Fifth, a*library\-level object*is needed because the main transition isℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}, not merely one tuple to another\.

#### The extended tuple\.

We therefore represent an editable skill record with seven fields

𝒮t=⟨Ct,πt,Tt,Rt,φt,νt,≺t⟩,\\mathcal\{S\}\_\{t\}\\;=\\;\\big\\langle\\,C\_\{t\},\\;\\pi\_\{t\},\\;T\_\{t\},\\;R\_\{t\},\\;\\varphi\_\{t\},\\;\\nu\_\{t\},\\;\\prec\_\{t\}\\,\\big\\rangle,\(2\)with the new components interpreted as skill\-record fields\. The tuple should not be read to mean that every skill owns an independent learning algorithm\. In most systems, the update policy is library\-level, while individual skill records expose the handles that policy can use: editable body, verification evidence, interface, and lineage\.

The*edit field*φt\\varphi\_\{t\}records how a candidate revision can be generated for this artifact under an edit instructionutu\_\{t\}\(chosen from the operator vocabulary introduced below in §[3\.3](https://arxiv.org/html/2607.10113#S3.SS3)\)\. In deterministic systems,φt:𝒮t×ut→𝒮t\+1\\varphi\_\{t\}:\\mathcal\{S\}\_\{t\}\\times u\_\{t\}\\to\\mathcal\{S\}\_\{t\+1\}maps directly to a successor skill\. In stochastic proposal systems such as mutation\-based search,φt\\varphi\_\{t\}is better read as a proposal kernelqt​\(𝒮′∣𝒮t,ut\)q\_\{t\}\(\\mathcal\{S\}^\{\\prime\}\\mid\\mathcal\{S\}\_\{t\},u\_\{t\}\), from which a realized candidate𝒮~t\+1\\tilde\{\\mathcal\{S\}\}\_\{t\+1\}is sampled before admission\. The edit component therefore need not guarantee improvement; the admission predicate below decides whether the realized candidate changes library state\. The*verification field*νt:𝒮t→\{0,1\}\\nu\_\{t\}:\\mathcal\{S\}\_\{t\}\\to\\\{0,1\\\}records the available admissibility handle—whether a proposed or edited skill can pass a quality gate before enteringℒt\+1\\mathcal\{L\}\_\{t\+1\}—and may be a unit test \(LATM\), a grounded rollout \(SkillWeaver,EvoSkill\), a meta\-agent inspection \(AgentFactory\), an ensemble judge \(Trace2Skill,AgentSkillOS\), a static analyzer \(SkillCraft\), a symbolic audit \(ASG\-SI\), or a Bayesian prior over future utility \(CODE\-SHARP\)\. The*lineage relation*≺t\\prec\_\{t\}records which version supersedes which; methods that support rollback, A/B maturity, or blast\-radius limits require an explicit≺\\prec\.

A static library is the limiting case in which no update trigger admits a library edit, soℒt\+1=ℒt\\mathcal\{L\}\_\{t\+1\}=\\mathcal\{L\}\_\{t\}except for exogenous releases\. An unconditional verifier with repeatedAddis not static; it is an unfiltered append\-only store\. The artifact type determines feasible\(φ,ν,≺\)\(\\varphi,\\nu,\\prec\)choices: executable skills support machine\-checkable gates, NL heuristics require indirect verification, parametric skills move edits onto a training timescale, and capability labels usually lack a well\-defined edit operator\. Cross\-skill operations such as composition, merging, and splitting are library\-level transitions, so we describe them next\.

### 3\.3Library dynamics:ℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}

A dynamic skill system is a library plus rules for how that library changes\. Write

ℒt=\{𝒮t\(1\),𝒮t\(2\),…,𝒮t\(Nt\)\}∪ℳt,\\mathcal\{L\}\_\{t\}\\;=\\;\\\{\\mathcal\{S\}^\{\(1\)\}\_\{t\},\\,\\mathcal\{S\}^\{\(2\)\}\_\{t\},\\,\\ldots,\\,\\mathcal\{S\}^\{\(N\_\{t\}\)\}\_\{t\}\\\}\\cup\\mathcal\{M\}\_\{t\},\(3\)whereℳt\\mathcal\{M\}\_\{t\}holds auxiliary metadata: invocation statistics, call graphs, maturity labels, verifier caches, and any cross\-skill edges \(prerequisite, composition, shared resources, author\)\. The library transition is then

ℒt\+1=𝒯​\(ℒt,τt,rt\)=Apply​\(u→t​\(τt,rt\),ℒt\),\\mathcal\{L\}\_\{t\+1\}\\;=\\;\\mathcal\{T\}\\\!\\big\(\\mathcal\{L\}\_\{t\},\\;\\tau\_\{t\},\\;r\_\{t\}\\big\)\\;=\\;\\mathrm\{Apply\}\\\!\\big\(\\vec\{u\}\_\{t\}\(\\tau\_\{t\},r\_\{t\}\),\\;\\mathcal\{L\}\_\{t\}\\big\),\(4\)whereτt\\tau\_\{t\}is an*evolution trigger*\(a timer, a task boundary, a failure event, a user edit\),rtr\_\{t\}is the*learning signal*the system uses to choose an edit \(task reward, natural\-language critique, self\-judgment, cross\-user aggregate, teacher signal\), andu→t\\vec\{u\}\_\{t\}is a vector of operator\-instruction pairs drawn from the ten\-element vocabulary

u→t⊆\{\\displaystyle\\vec\{u\}\_\{t\}\\subseteq\\\{Add,Refine,Merge,Split,Prune,Distill,Abstract,\\displaystyle\\textsc\{Add\},\\textsc\{Refine\},\\textsc\{Merge\},\\textsc\{Split\},\\textsc\{Prune\},\\textsc\{Distill\},\\textsc\{Abstract\},Compose,Rewrite,Rerank\}×Instruction\.\\displaystyle\\textsc\{Compose\},\\textsc\{Rewrite\},\\textsc\{Rerank\}\\\}\\times\\mathrm\{Instruction\}\.The ten operators have fixed meanings throughout the paper:Addinserts a skill;Refineedits content without changing the interface;Mergecombines skills;Splitfactors one skill into components;Pruneremoves or quarantines;Distillcompresses trajectories into a skill;Abstractlifts a concrete procedure to a template;Composechains skills into a composite;Rewritechanges the body and possibly the interface; andRerankchanges retrieval priors without changing content\. Representative instances includeVoyagerandSAGEforAdd,Memento,PSN,SkillOpt, andSkillGradforRefine,Trace2SkillandAutoRefineforMerge,SkillXforSplit,Wild\-SkillsandClawSafety\(Weiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib56)\)forPrune,CASCADEandMUSEforDistill,CUA\-Skill,CoEvoSkills, andSkillGenforAbstract,SkillCraft,SkillOrchestra, andHASPforCompose,EvolveRandEmbodiSkillforRewrite, andSkillRouterandSkillsInjectorforRerank\. Almost no method implements all ten; the supported subset is one of the clearest taxonomic fingerprints\.

#### Verification as library\-level gate\.

Althoughν\\nuis indexed per skill record in Equation[2](https://arxiv.org/html/2607.10113#S3.E2), verification acts at*admission*time: an edit yields a candidate𝒮∗\\mathcal\{S\}^\{\*\}, and𝒮∗\\mathcal\{S\}^\{\*\}entersℒt\+1\\mathcal\{L\}\_\{t\+1\}only if the library\-level admission policy accepts the candidate using the available verification handle\. This edit\-versus\-admission distinction underlies the verification architecture of Section[7](https://arxiv.org/html/2607.10113#S7)and the R1 pattern in Section[9](https://arxiv.org/html/2607.10113#S9)\.

#### Two timescales\.

The triggerτt\\tau\_\{t\}can be a per\-step failure \(SELAUR\), per\-task retrospective pass \(SAGE,ERL\), periodic maintenance cycle \(AutoRefine\), or release decision \(AgentDevel\)\. Parametric systems such asMetaClaw,SKILL0, andLSEadd a slow loop in which many fast\-loop library updates are distilled into weights or adapters\.

#### Scope\.

Throughout the rest of the paper, “dynamic skill” refers to a skill𝒮t\\mathcal\{S\}\_\{t\}with an editable body, an admission/verification handle, or lineage metadata, or to a libraryℒt\\mathcal\{L\}\_\{t\}that evolves under a non\-empty update vectoru→t\\vec\{u\}\_\{t\}\. A*static*library is the special caseu→t≡∅\\vec\{u\}\_\{t\}\\equiv\\varnothing\. The survey addresses the dynamic case; static\-library methods are included only as baselines or as infrastructure substrates on which dynamic methods are built\.

### 3\.4Worked instantiation: AutoRefine as a library transition

The notation is intended to describe concrete system behavior, not just to name components\. Consider anAutoRefine\-style maintenance cycle\(Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25)\)\. A skill document in the current library can be written as

𝒮t\(i\)=⟨Ct,πt,Tt,Rt,φt,νt,≺t⟩,\\mathcal\{S\}^\{\(i\)\}\_\{t\}=\\langle C\_\{t\},\\pi\_\{t\},T\_\{t\},R\_\{t\},\\varphi\_\{t\},\\nu\_\{t\},\\prec\_\{t\}\\rangle,whereCtC\_\{t\}is the task or state description under which the skill should be retrieved,πt\\pi\_\{t\}is the SKILL\.md instruction body plus optional executable helper,TtT\_\{t\}is the success/failure or budget condition observed during use, andRtR\_\{t\}is the call signature or natural\-language invocation handle\. After a trajectory exposes a failure or redundancy, the trigger isτt=TaskEnd\\tau\_\{t\}=\\textsc\{TaskEnd\}orPeriodic, and the signalrtr\_\{t\}is a critique, execution trace, or downstream utility measurement\. The update rule chooses an operator vector such as

u→t=\{\\displaystyle\\vec\{u\}\_\{t\}=\\\{\(Refine,patch ambiguous step\),\\displaystyle\(\\textsc\{Refine\},\\text\{patch ambiguous step\}\),\(Merge,combine duplicate routines\),\\displaystyle\(\\textsc\{Merge\},\\text\{combine duplicate routines\}\),\(Prune,quarantine low\-utility skill\)\}\.\\displaystyle\(\\textsc\{Prune\},\\text\{quarantine low\-utility skill\}\)\\\}\.Each realized edit yields a candidate𝒮∗\\mathcal\{S\}^\{\\ast\}\. The admission gate evaluatesνt​\(𝒮∗\)\\nu\_\{t\}\(\\mathcal\{S\}^\{\\ast\}\)using the method’s judge, execution feedback, or consistency checks\. If the candidate passes,𝒮∗\\mathcal\{S\}^\{\\ast\}entersℒt\+1\\mathcal\{L\}\_\{t\+1\}and≺t\+1\\prec\_\{t\+1\}records that it supersedes or merges earlier artifacts; if it fails,ℒt\+1\\mathcal\{L\}\_\{t\+1\}retains the prior version or marks the candidate for later review\. In this example, the transition in Equation[4](https://arxiv.org/html/2607.10113#S3.E4)is not a single append operation: it is a gated composition ofRefine,Merge, andPruneover a versioned artifact store\.

## 4Why Static Skill Libraries Fail

The record schema of §[3](https://arxiv.org/html/2607.10113#S3)treats the skill library as a time\-indexed objectℒt\\mathcal\{L\}\_\{t\}because the static alternative fails in recurring, connected ways\. Static libraries are useful when the task distribution, tool surface, verifier, and authoring assumptions remain stable\. The surveyed literature shows that long\-lived agent deployments rarely satisfy those conditions\. The failure is not one defect but a lifecycle collapse: authoring, verification, retrieval, provenance, and adaptation are all forced into a one\-time design decision\.

The first pressure is economic\. Static skills require up\-front human authoring, yet deployment value is only observed after the skill has been used in context\. Deployment\-facing papers such asSkillClaw,AutoSkill, andAgentSkillOSdescribe useful SKILL\.md authoring as labor\-intensive, andSoK\-Skillsobserves that library quality is bounded by the weakest authors\. The issue is not simply that documentation is costly; it is that a static library pays the cost per skill before knowing which skills will matter\. Dynamic systems shift part of that cost to write\-timeAbstractandDistill, so trajectory evidence decides what should become reusable\.

The second pressure is that correctness is not stationary\. A static library freezes the author’s implicit verifier at authoring time, but tasks, tools, base models, runtimes, and safety constraints drift\.PSN’s refactor detectors depend on recent transition history;AgentSkillOS’s capability\-tree audits assume a current tool surface;CODE\-SHARP’s execution verifier assumes a compatible runtime\. Once those assumptions move, a skill can remain syntactically valid while becoming operationally wrong\. Dynamic systems makeν\\nure\-runnable and allow the admission gate to remove, quarantine, or demote skills whose verifier has drifted\.

The third pressure appears as the library grows: selection becomes harder than storage\. Controlled size sweeps show that flat retrieval can degrade in the moderate\-library\-size regime, often around tens to hundreds of skills\.Single\-Agent\-Skillsgives the cleanest controlled curve, whileWild\-Skills,SkillRouter, andAgentSkillOSshow related degradation under realistic distractors, 80K\-scale routing, and large flat stores\(Li,[2026](https://arxiv.org/html/2607.10113#bib.bib52); Liuet al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib27); Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43); Liet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib42)\)\. The exact threshold varies, but the mechanism is stable: adding skills eventually adds distractors faster than utility\. A static system can cap the library or redesign retrieval, but it has no native way to ask whether old skills should be merged, pruned, or reranked\. Dynamic systems add those maintenance operators:Prune,Merge, andRerank\.

The fourth pressure is provenance\. Ordinary version control records who edited a file and when, but not which trajectory, verifier, or deployment signal justified admission\. That gap matters when a skill regresses \(AutoRefine,AgentSkillOS\), when admission must be re\-run against a new verifier, or when a cross\-user skill must be attributed, redacted, or rolled back \(SkillClaw,AutoSkill\)\.PSN’s rollback gate makes the point concrete: tentative refactors are reverted if success on three recent tasks drops by more than20%20\\%\. The lineage relation≺\\precis the minimal structural addition that makes rollback, re\-admission, and provenance tractable\.

Finally, deployment itself is non\-stationary\. A static library is a snapshot of the authoring distribution, but real task mixes shift\. The surveyed evidence adds two important shapes to this familiar observation: weaker backbones gain disproportionately from dynamic skills, and focused libraries become stale when the task mix changes\. The response is not merely to refresh the library occasionally; it is to specify an evolution triggerτ\\tauand a fast\-loop clock that decide when evidence is allowed to alter the library\.

These failures explain why the added record fields are not cosmetic\. Authoring cost points toφ\\varphi; verifier drift points to re\-runnableν\\nu; retrieval pollution points to maintenance operators; attribution loss points to≺\\prec; and task non\-stationarity points to explicit triggers\. Table[2](https://arxiv.org/html/2607.10113#S5.T2)reads the same mapping from the architecture side, and TableLABEL:tab:mastermaps methods to the lifecycle gaps they address\.

## 5Dynamic Skill Systems as Lifecycle\-Managed Stores

Section[4](https://arxiv.org/html/2607.10113#S4)gives the negative case: static skill libraries fail because they make authoring, verification, retrieval, provenance, and adaptation one\-time decisions\. The positive object is therefore not just a larger skill library\. It is a controlled state machine over an evolving store of artifacts\. A dynamic skill system observes interaction evidence, proposes a skill or library edit, verifies the candidate, admits it into a storage topology, retrieves or composes it at future invocation time, maintains it as the library ages, and records enough provenance to support rollback, transfer, and governance\.

This section defines that reference architecture\. The next section uses it as a taxonomy for the surveyed papers; Section[7](https://arxiv.org/html/2607.10113#S7)then analyzes the implementation choices that make particular lifecycle stages possible\. Keeping these roles separate is important: the lifecycle is the*architecture*; the taxonomy is the*classification of systems*; the mechanism design space is the*choice of operators, verifiers, and clocks*\.

### 5\.1Reference Architecture

Figure[1](https://arxiv.org/html/2607.10113#S5.F1)gives the high\-level architecture; Table[2](https://arxiv.org/html/2607.10113#S5.T2)records the corresponding design questions, operators, evidence, and failure modes\. We decompose a dynamic skill system into eight recurring stages\.*Evidence acquisition*decides what observation can justify a library change: a trajectory, reward, failure trace, user edit, cross\-user signal, or external resource\.*Proposal*converts that evidence into a candidate artifact or edit\.*Verification and admission*decides whether the candidate is allowed to enter the library, and whether it enters as mature, tentative, quarantined, or rejected\.*Organization and storage*assigns the admitted artifact to a flat index, hierarchy, DAG, invocation graph, ontology, dual memory/skill store, or parametric subspace\.*Retrieval and composition*decides which artifacts influence a future action\.*Maintenance and repair*keeps the library compact and correct by pruning, merging, splitting, reranking, refining, or rewriting artifacts\.*Distillation and portability*moves procedural knowledge between external artifacts, model weights, agents, users, or task domains\.*Governance and provenance*records lineage, authorship, safety checks, release state, and rollback handles; in implementations this usually means audit logs over admitted candidates, verifier decisions, operator edits, and rollback events\.

The stage order should not be read as a waterfall\. Many systems loop between proposal and verification, perform retrieval before maintenance, or run distillation only periodically\. The point is that each stage asks a different design question\. A system that performs strong retrieval has not necessarily solved admission; a system that frequently uses a skill has not necessarily shown utility; a system that can add skills has not necessarily learned how to remove or repair them\.

![Refer to caption](https://arxiv.org/html/2607.10113v1/figures/dynamic-skill-lifecycle.png)Figure 1:Dynamic skill systems as lifecycle\-managed artifact stores\.Interaction evidence drives proposal and verification; admitted artifacts enter an evolving skill store, where retrieval and execution create further evidence\. Maintenance repairs, prunes, merges, or reranks the store over time\. Governance and provenance wrap the lifecycle through shields, audit logs, lineage records, and rollback handles, while distillation and portability form a slower side loop\. The illustration compresses the textual eight\-stage lifecycle: organization/storage is represented by the central store, execution is the action point that produces new evidence, and governance/distillation are wrapper or side\-loop functions rather than extra linear stages\.Table 2:Lifecycle architecture for dynamic skill systems\.The table is not a method ranking; it identifies the recurring stages at which a system must make design commitments\. Later tables specialize this architecture into system families, verification architectures, evidence\-graded patterns, and open problems\.
### 5\.2The Admission Boundary

The most important boundary in the lifecycle is between*candidate*artifacts and*admitted*library state\. Dynamic systems can generate many plausible skills cheaply, but the library only improves when the admission gate filters them with an appropriate verifier\. Execution\-grounded systems such asSkillWeaver,EvoSkill, andSkillFoundryplace the gate close to write time; maintenance\-heavy systems such asAutoRefineandAgentSkillOSre\-run the gate as the library ages; rollback systems such asPSNtreat admission as tentative until recent\-task utility remains stable\. This boundary explains why verification is not a side module\. It is the selection mechanism that turns skill generation into library learning\.

Admission also determines what provenance must be stored\. If a skill was admitted because of a trajectory, validator, cross\-user signal, or release audit, the system must retain that justification so that future maintenance can demote, revise, or remove the artifact\. Without this lineage, dynamic skills become append\-only memories with a more polished file format\.

### 5\.3What Counts as Dynamic

We use “dynamic” for systems with a non\-trivial library transitionℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}, not for any system that retrieves a skill at inference time\. Retrieval changes the context; dynamic update changes the store\. A method can therefore be skill\-using without being dynamic, and it can be dynamic in a narrow way if it implements only one transition such asAddafter each task\. Lifecycle maturity increases as systems add admission, maintenance, lineage, governance, and two\-timescale consolidation\.

This distinction sets up the taxonomy in Section[6](https://arxiv.org/html/2607.10113#S6)\. The lifecycle table above defines the design commitments a complete dynamic skill system must make; the taxonomy asks which subsets of those commitments each paper actually implements\.

## 6A Lifecycle Taxonomy of Dynamic Skill Systems

The taxonomy asks which lifecycle configuration a system instantiates\. We use two levels: lifecycle families for the conceptual map, and seven coding fields in the master coding sheet \(TableLABEL:tab:master, Appendix[A](https://arxiv.org/html/2607.10113#A1)\) for auditability\.

2023\+2\+22early librariesVoyager,LATM2024\+0\+02no primarycounted papers2025\+19\+1921transition waveSkillWeaver,SkillFlow2026\+103\+103lifecycle ecosystemmethods, benchmarks; infrastructure, safety124May 31, 2026 cutoff2026 clusterDynamic librariesBenchmarksInfrastructureSafety/governance

Figure 2:Temporal structure of the dynamic\-skills audit set\.Counts summarize the 124 modern papers covered by the survey and exclude only older classical\-options and HRL background anchors; boundary/context papers are included for scope but are not treated as primary causal evidence in the evidence\-graded patterns\. The figure reports annual additions and cumulative coverage through the May 31, 2026 cutoff\. Representative work names are milestones rather than an exhaustive bibliography\.### 6\.1Primary Families

Table[3](https://arxiv.org/html/2607.10113#S6.T3)separates systems that are often conflated\. Retrospective lesson systems and PPVH systems both update after a task, but only PPVH has execution\-grounded admission\. Skill\-aware RL and two\-timescale systems both use reward, but only the latter separates fast external edits from slow parametric internalization\. Cross\-user transfer and registry\-scale infrastructure both handle many skills, but one concerns distributed authorship and the other routing/storage\. Benchmarks such asSkillFlow\-BenchandWild\-Skillsare included because they expose lifecycle behavior method papers often hide\.

Table 3:Primary lifecycle taxonomy of dynamic skill systems\.Rows are families, not rankings\. The table compresses each family into its dominant lifecycle stages, typical operator footprint, assurance bottleneck, and residual failure mode\. This makes the comparison sharper than a per\-paper list: families differ mainly in which parts ofℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}they make cheap, verified, or governable\.
### 6\.2Seven Coding Fields and Three Couplings

We code each representative system along seven fields:*artifact type*\(code, NL lesson, SKILL\.md package, memory trace, graph/workflow, or weight delta\),*update locus*,*evolution trigger*,*operator repertoire*,*learning signal*,*storage topology*, and*model portability*\. These are not claimed to be orthogonal basis dimensions\. They are audit fields whose couplings are often the point\.

Three couplings matter most\.*Artifact–verifier coupling*: executable code supports tests and rollouts, while NL lessons and capability labels rely on weaker downstream or judge\-based checks\.*Storage–maintenance coupling*: flat stores makeAddcheap but makeMergeandPrunecostly; graphs and hierarchies expose structure but require more careful rollback\.*Trigger–signal coupling*: task\-end retrospection naturally supplies critique, RL loops supply reward, user edits supply ownership constraints, and deployment registries supply aggregate utility\. This dependence is why Table[3](https://arxiv.org/html/2607.10113#S6.T3)is the conceptual taxonomy and TableLABEL:tab:masteris the coding sheet\. Figure[3](https://arxiv.org/html/2607.10113#S6.F3)gives the corresponding visual summary: families differ less in whether they “use skills” than in which lifecycle stages they actually cover\.

![Refer to caption](https://arxiv.org/html/2607.10113v1/figures/lifecycle-coverage-heatmap.png)Figure 3:Lifecycle coverage across dynamic\-skill system families\.Cell intensity summarizes how centrally each family implements or evaluates a lifecycle stage in the coding sheet\. The figure is a synthesis device rather than a ranking: executable libraries concentrate around proposal, verification, admission, storage, and retrieval; graph systems concentrate around storage, retrieval, and maintenance; two\-timescale systems concentrate around distillation; benchmarks expose evidence and verification protocols; registries emphasize storage and retrieval; and safety/governance systems concentrate on admission, verification, and provenance\.
### 6\.3Master Coding Table

TableLABEL:tab:masterin Appendix[A](https://arxiv.org/html/2607.10113#A1)instantiates the seven coding fields for representative systems\. The “headline” column is factual rather than evaluative so that the table remains a map, not a comment collection\. We place the full coding sheet in the appendix because it is an audit artifact; the main text uses Table[3](https://arxiv.org/html/2607.10113#S6.T3)and Figure[3](https://arxiv.org/html/2607.10113#S6.F3)for synthesis\.

### 6\.4What the Taxonomy Reveals

Three diagonals matter most\. Artifact and storage co\-vary; trigger and signal co\-vary; and lifecycle maturity is visible from the operator set\. The newest infrastructure and safety papers add a fourth diagonal: once skills are packaged for registries, scanners, package managers, and repository\-context checks become part of the same taxonomy as routing and maintenance\. These diagonals motivate the mechanism synthesis of §[7](https://arxiv.org/html/2607.10113#S7): operator vocabulary, verification architecture, and fast–slow adaptation\. They should be read as coded patterns in a heterogeneous corpus, not as causal conclusions\.

## 7Mechanisms: How Dynamic Skill Stores Improve

Sections[5](https://arxiv.org/html/2607.10113#S5)–[6](https://arxiv.org/html/2607.10113#S6)define the object of study and classify the corpus\. This section asks a narrower question: what mechanism turns experience into a better skill store? Across the literature, four choices are load\-bearing\. A system must decide which edits it can make, how candidate edits are admitted, how the resulting store is organized and retrieved, and when external skills should be consolidated into slower parametric or shared stores\. These choices are coupled: a wide edit repertoire requires stronger verification; a large store requires stronger routing and maintenance; and distillation is useful only when the fast\-loop library is already selective enough to provide a clean training signal\.

### 7\.1Edit repertoire: expansion, compression, and refactoring

The operator vocabulary of Equation[4](https://arxiv.org/html/2607.10113#S3.E4)is most useful as a diagnostic for library maturity\. The simplest systems are*expansion\-only*: they add or refine skills after a task, as inVoyager,LATM,SAGE,ERL, andRetroAgent\(Wanget al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib1); Caiet al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib2); Wanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib3); Allardet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib7); Zhanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib8)\)\. Stronger executable variants already pairAddwith abstraction or composition:ASIturns successful web trajectories into verified program actions, whileWebXSkillandSkillDroidcompile trajectory fragments into reusable web or mobile GUI skill templates\(Wanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib100);[2026g](https://arxiv.org/html/2607.10113#bib.bib101); Chenet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib103)\)\. This is enough for short horizons, but it makes the library monotone unless later stages remove, merge, or repair stale artifacts\.

Mature systems add a second class of operators that compress or discipline the store\.Prune,Merge, andRerankappear inWild\-Skills,SkillRouter,AutoRefine,Trace2Skill, andAgentSkillOS;SkillOps\(Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\)makes the same point explicit as library\-time technical\-debt management through merge, repair, retire, validator insertion, and adapter insertion\. The recurring lesson is that a dynamic library needs a negative operator once distractor load matters\. Growth alone is not learning\. The mechanism\-level question is therefore not only how a system writes skills, but how it removes, merges, or demotes them when later evidence says they are unhelpful\.

The late\-May 2026 wave adds a sharper specialization:*skill optimization*\.SkillOpt\(Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128)\)treats a skill document as external state optimized by bounded add/delete/replace edits under held\-out validation;SkillGrad\(Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130)\)casts failed and contrastive\-successful executions as text gradients plus momentum over a structured skill package; andSkillGen\(Maet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib119)\)verifies a synthesized skill by its net interventional effect, including both repaired failures and induced regressions\. These papers do not introduce a new lifecycle stage, but they make the write\-time transitionℒt→ℒt\+1\\mathcal\{L\}\_\{t\}\\to\\mathcal\{L\}\_\{t\+1\}look less like reflection and more like optimization over an editable artifact\.

A third class changes structure rather than content\.PSNapplies rollback\-gated refactors over an invocation graph;CODE\-SHARPmutates a prerequisite DAG;ContractSkillcompiles loose web skills into contracts with local patch sites;SkillXsplits hierarchies;CoEvoSkillsabstracts multi\-fileSKILL\.mdpackages; andBilevel\-MCTSandSkillMOOoptimize package structure and pass\-rate/cost trade\-offs\(Shiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib17); Bornemannet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib6); Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102); Wanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib21); Zhanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib16); Huanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib90); Gonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib76)\)\. Structural edits are powerful because they can change reuse pathways, not just local skill text, but they also require lineage≺\\precand rollback because a bad rewrite can invalidate many downstream invocations\.

Parametric methods occupy a different regime\. Their visible operator set often collapses to\{Distill,Abstract\}\\\{\\textsc\{Distill\},\\textsc\{Abstract\}\\\}:Refinebecomes more training,Mergebecomes weight\-space interpolation, andPruneis no longer a surgical library edit\. This is whySkillsCrafter,MetaClaw,SKILL0, andK2\-Agentare best read as two\-timescale systems rather than ordinary library editors\(Wanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib33); Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35); Luet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib31); Wuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib34)\)\.

### 7\.2Admission and verification are the selection mechanism

Verification is not an auxiliary module; it is the selection pressure that decides which generated artifacts become library state\. Table[4](https://arxiv.org/html/2607.10113#S7.T4)organizes the main verifier forms\. Execution verifiers catch runnable failures through tests, contracts, simulators, or environment rollouts \(SkillWeaver,ASI,WebXSkill,ContractSkill,SkillDroid,EvoSkill,LIVE\-SWE,SkillFoundry,SkillCraft\)\. Judge\-LLM verifiers assess semantic quality, pairwise preference, failure diagnosis, or release readiness \(Trace2Skill,CoEvoSkills,AutoRefine,AgentSkillOS,SkillForge,MedSkillAudit\)\. Rollback and audited\-graph verifiers turn verification into a state\-transition check, as inPSNandASG\-SI\. Utility\-based verifiers defer part of the decision to downstream use, as inWild\-Skills,SWE\-Skills\-Bench,EffiSkill,SkillGen,SkillOpt, andSkillMaster\.

The important distinction is what each verifier can observe\. Execution gates are precise but narrow: passing one contract does not prove broad transfer\. Judge gates are broader but vulnerable to evaluator drift and rubric hacking\. Rollback gates directly measure behavioral regression, but only on the probe set\. Utility gates are cheap and deployment\-realistic, but they may admit harmful or misleading skills before enough evidence accumulates\. These limitations explain why high\-capacity systems increasingly use staged admission: a candidate can be tentative, quarantined, promoted, demoted, or rolled back rather than simply accepted or rejected\.

- aTiming codes:W= write\-time,I= invocation\-time,M= maintenance\-time \(periodic sweep over the library\)\.
- bAdmission policies:*hard*thresholding = admit iff verifier returns success;*Pareto*= admit iff candidate dominates existing skills on a vector of criteria;*maturity\-gated*= candidate enters as trial, promoted after usage budget;*graph integrity*= admit iff graph\-level audit predicates hold \(ASG\-SI\);*Δ\\Delta\-success≤20%\\leq 20\\%*= tentative admission reverted if task\-success rate on recent tasks drops by more than 20% \(PSN\);*stability\-gated*= admit iff fast\-loop performance variance is below threshold;*utility threshold*= defer admission, evict below utility floor\.
- cExecution verification is “narrow” because a passing unit test or rollout does not imply generalisation; this is partially why execution\-verified methods cap their operator repertoire at\{Add,Refine,Prune\}\\\{\\textsc\{Add\},\\textsc\{Refine\},\\textsc\{Prune\}\\\}\(§[7\.2](https://arxiv.org/html/2607.10113#S7.SS2)\)\.
- dAgentFactory’s verifier is a meta\-agent inspection of proposed subagents against a design specification; we classify it here as a judge\-LLM variant \(rather than as execution\) because it inspects the proposal rather than running it against unit\-test\-style contracts\.
- eAgentic\-Proposinguses a three\-LLM\-judge ensemble with majority voting\.Agent0andASG\-SIare driven by curriculum\-reward and audit\-gated reward respectively; they are not verifier ensembles in the sense used in this row and therefore do not appear here\.

Table 4:Verification architectures in the surveyed dynamic\-skill methods\.Three axes organize the space: verifier*form*, verification*timing*, and*admission policy*\. The two rightmost columns record the qualitative cost per candidate skill and the coverage over the class of defects each form can catch\.
### 7\.3Storage and retrieval control the scaling regime

After admission, the bottleneck shifts from writing good skills to finding the right skills without importing distractors\. Flat retrieval is attractive because it is simple, but the surveyed scaling studies repeatedly show a moderate\-library\-size drop\.Single\-Agent\-Skillsreports a sharp decline beyond roughly 64–128 skills when many agent skills are compiled into a single\-agent library;Wild\-Skillsshows that realistic distractors can erase apparent gains from curated skills;SkillRoutershows that 80K\-scale registries require full\-text retrieve\-and\-rerank rather than metadata\-only selection;SRAshows that retrieval, incorporation, and downstream utility must be evaluated separately because agents may load skills at similar rates regardless of whether the gold skill is present or needed; andSkillsInjector\(Liet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib131)\)shows that even a fixed library needs adaptive selection, budgeting, and set\-aware description rendering because static all\-injection can degrade performance\(Li,[2026](https://arxiv.org/html/2607.10113#bib.bib52); Liuet al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib27); Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43); Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)\.

The response is to make storage reflect reusable structure\. Hierarchies and capability trees support audit and specialization \(AgentSkillOS,SkillX,Corpus2Skill,Uni\-Skill\); DAGs encode prerequisites or typed composition \(CODE\-SHARP,GraSP\); invocation and relation graphs support refactoring, dependency\-aware retrieval, and provenance \(PSN,Graph of Skills,SkillGraph,WebXSkill,ASG\-SI\)\(Nieet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib98)\); typed ecosystem graphs expose dependency, compatibility, redundancy, and alternative edges for library health diagnosis \(SkillOps\); structured intermediate representations make retrieval and audit less dependent on raw prose \(SSL\)\(Lianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib112)\); compilation and boundary extraction make runtime interfaces smaller \(SkillSmith\)\(Xuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib122)\); ontologies and registries support portability and package\-level governance \(SkillNet,Skilldex,SkillsVote\)\(Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125)\)\. The mechanism\-level insight is that retrieval quality is partly an indexing problem and partly a maintenance problem\. Once a store grows, storage topology, reranking, pruning, and provenance become one system\.

### 7\.4Update clocks separate adaptation from consolidation

Many recent systems separate a fast external loop from a slow internal loop\. The fast loop edits files, procedures, code snippets, or graph nodes after tasks; the slow loop distills selected behavior into adapters, weights, shared libraries, or compact package formats\. This separation lets the agent learn quickly without paying the cost or risk of continuous model updates, while still allowing stable skills to become cheaper to invoke later\.

Table[5](https://arxiv.org/html/2607.10113#S7.T5)compares the recurring decouplings\. Some systems share the artifact between loops \(MetaClaw,K2\-Agent\); others share reward or critique signals \(SAGE,AgentEvolver,Tool\-R0,COS\-PLAY\); others share curricula or verifiers \(Agent0,SCALAR,ASG\-SI\)\. The key design variable is promotion: prevalence is cheap, reward is task\-grounded, coverage favors behavioral diversity, and stability is closest to measuring whether distillation will preserve rather than corrupt a skill\.

Two\-timescale adaptation is not automatically superior\. A noisy fast\-loop library gives the slow loop noisy targets, so distillation can compress mistakes as well as discoveries\. Conversely, an over\-conservative fast loop leaves useful knowledge external and expensive\. The open variable is the consolidation schedule: when to distill, what to distill, and whether the slow\-loop result should replace, augment, or merely rerank the external library\.

- aDecoupling mode \(§[7\.4](https://arxiv.org/html/2607.10113#S7.SS4)\):SA= shared\-artifact \(fast and slow loops edit and distill the same textual skill\);SS= shared\-signal \(both loops consume the same reward / critique stream at different cadences\);SC= shared\-curriculum \(fast loop generates tasks for the slow loop\);SV= shared\-verifier \(same verifier gates both loops at different stringency\)\.
- bSAGE’s fast artifact is an executable Python skill function updated under a skill\-integrated GRPO reward; the slow loop augments this artifact rather than replacing it, and we flag the row because it is the onlySSmethod whose fast\-loop artifact is code rather than a natural\-language heuristic\.
- cMementois listed as a fast\-only baseline for contrast; it has no slow loop and therefore no decoupling mode applies\.
- dLSEtrains a 4B edit policy that emits context\-level edits as a learned action; there is no separate fast/slow decomposition because the learned editor*is*the fast loop and its weights are the only parametric artifact, so the four modes do not apply\.
- eCo\-Evolvingruns alternating DPO over a hard\-negative trajectory pool; critiques and policy updates live in a single optimization loop, with no separate slow\-loop distillation over a persistent skill library\.
- fAgentic\-Proposingpairs a fast loop that proposes*problems*\(not library edits\) with an RL loop that updates the policy against verifier\-gated rewards; because the fast\-loop output is not a skill artifact, the two\-timescale decoupling modes above do not apply\.

Table 5:Two\-timescale decoupling modes in methods with both a fast in\-context library and a slow parametric loop\.Rows are keyed by method and grouped by the four decoupling modes the literature converges on\. The “Promoted on”, “Slow trigger”, and “Slow vs\. fast” columns characterize what moves between loops, when, and whether the slow\-loop output replaces or augments the fast\-loop library\.
### 7\.5Mechanism\-level synthesis

The mechanism picture is compact\. Dynamic skill systems improve when expansion is paired with compression, admission is paired with re\-verification, retrieval is paired with structure, and fast editing is paired with slower consolidation\. The same four requirements explain why apparently different systems occupy coherent regimes: executable\-skill systems emphasize execution gates and small edit repertoires; natural\-language lesson systems emphasize cheap proposal and later maintenance; graph and hierarchy systems trade storage complexity for retrieval and refactoring; and parametric systems trade editability for amortized inference\. The strongest systems are not those with the largest libraries, but those that make the library easier to verify, route, repair, and consolidate over time\.

## 8Evaluation

A benchmark that reports only endpoint task success hides the phenomena that define dynamic skills: skill inflation, incorrect\-skill drift, maintenance\-off collapse, retrieval degradation as flat libraries grow, and usage without utility\. This section audits evaluation through a lifecycle lens: how libraries are created, repaired, routed, compacted, and governed over time\.

### 8\.1The Benchmark Landscape

The surveyed papers report on four partly\-overlapping benchmark clusters\. The first is the*agent\-task*cluster: WebArena, VisualWebArena, Mind2Web, AgentBench, ST\-Bench, SWE\-bench\-style command\-line environments, and related terminal or software\-engineering suites\. These benchmarks evaluate an agent end\-to\-end on held\-out tasks; dynamic\-skill papers often inherit the benchmark and report a single success rate\. The second is the*code\-execution*cluster: HumanEval, MBPP, LiveCodeBench, APPS, and software\-engineering task suites with intrinsic execution verification\. These benchmarks are why code\-skill and skill\-aware RL papers can afford stronger admission gates\. The third is the*reasoning*cluster: AIME, MATH, GPQA, and HLE; this is whereAgentic\-ProposingandMementoreport some headline results, and where verifier quality is often load\-bearing because ground truth is unambiguous\.

The fourth cluster is most specific to this survey:*skill\-lifecycle evaluation*\.SkillsBench\(Liet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib65)\)measures skill usefulness across tasks,SWE\-Skills\-Bench\(Hanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib104)\)isolates the marginal value of public SWE skills under deterministic tests,Wild\-Skills\(Liuet al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib27)\)stresses retrieval realism under distractors and 34K\-candidate retrieval,Single\-Agent\-Skills\(Li,[2026](https://arxiv.org/html/2607.10113#bib.bib52)\)gives the clearest controlled size sweep for 64–128\-skill flat\-retrieval degradation,SkillRouter\(Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43)\)evaluates full\-text routing over an 80K skill pool, andSRA\(Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)decomposes large\-corpus skill augmentation into retrieval, incorporation, and end\-task utility over a 26,262\-skill corpus\.SkillFlow\-Bench\(Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48)\)is the most lifecycle\-aligned benchmark: 166 runnable tasks across 20 DAEF\-structured families, empty initial family libraries, JSON skill patches from trajectories, and reports of completion, turns, cost, output tokens, final skill count, and skill\-use rate\. Its Table 1 gives the key finding that skill use is not skill utility: Claude Opus 4\.6 improves from62\.65%62\.65\\%to71\.08%71\.08\\%, while Kimi K2\.5 gains only\+0\.60\+0\.60points despite66\.87%66\.87\\%skill use and GPT 5\.3 Codex regresses by6\.026\.02points\.

SkillLearnBench\(Zhonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib82)\)asks whether agents can continually generate useful skills, not merely use supplied ones, and finds a large gap to human performance across 20 verified skill\-dependent tasks\.Raw\-Experience\(Huanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib127)\)broadens this into a lifecycle study over experience generation, skill extraction, and skill consumption, finding that model\-generated skills help on average but can transfer negatively and that a strong extractor need not be a strong consumer\.SkillGen\(Maet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib119)\)gives the interventional version of the same evaluation idea: candidate skills are selected by net effect, counting both repaired failures and new regressions\.SWE\-Skills\-Benchgives the complementary negative result for supplied skills: across 49 public SWE skills and about 565 tasks, average pass\-rate gain is only\+1\.2%\+1\.2\\%, 39 skills yield no improvement, and three regress\.HarmfulSkillBench\(Jianget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib80)\)is the safety analogue, separating harmful skill presence from explicit and implicit invocation\.MedSkillAudit\(Houet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib93)\)adds a domain\-release protocol: evaluate the reusable skill artifact itself for release readiness\.

Table 6:Benchmark coverage over dynamic\-skill lifecycle dimensions\.The table separates benchmark roles rather than ranking benchmarks\. “Self\-gen\.” means the benchmark evaluates skills produced by the agent; “Revision” means skills can be patched or repaired over time; “Trajectory” means the protocol exposes a time\-ordered library or task sequence; “Usage utility” means the benchmark can distinguish reading/calling a skill from actually improving task outcome\.
### 8\.2Reported Metric Families

Four metric families dominate\.*Terminal success rate*or pass@kkremains the default: evaluate after training or after library construction and report a single number\. This metric is useful but incomplete because it hides library\-size effects, focus effects, and maintenance effects\.*Sample efficiency*or wall\-clock\-to\-threshold metrics \(EvoSkill,SAGE,K2\-Agent,SkillDroid,SkillOpt,SkillGrad\) report how quickly a method reaches a target success rate or reduces repeated\-use cost; these are better for surfacing the weaker\-backbone and repeated\-use advantages that dynamic skills often show\.*Library\-size and retrieval stress tests*\(Single\-Agent\-Skills,Wild\-Skills,SkillRouter,Graph of Skills,SRA,SkillsInjector\) expose scaling behavior, but only a minority of papers report them\.*Lifecycle metrics*, newly visible inSkillFlow\-Bench,SkillLearnBench,SWE\-Skills\-Bench,Raw\-Experience,SkillsVote, andSRA, report final skill count, file\-kind composition, skill\-use rate, generated\-skill quality, retrieval\-versus\-incorporation gaps, and the gap between skill usage and task improvement\.

XSkill\(Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12)\)adds a multimodal metric design: Average@44and Pass@44over repeated rollouts, plus ablations over the skill stream, experience stream, and knowledge managers\. It is not a full library\-trajectory protocol, but it separates average rollout quality from best\-of\-four exploration and shows that transfer can raise Pass@44while lowering Average@44on weaker open\-source backbones\.

Two quantities remain under\-reported\. The first is*operator velocity*: counts ofAdd,Refine,Merge,Prune, andRerankper task or per dollar\. Without this quantity, the velocity–soundness tradeoff betweenPSN,CODE\-SHARP,SkillMOO, andBilevel\-MCTScannot be compared quantitatively\. The second is*repair quality*: whether a later skill patch actually corrects a faulty abstraction\.SkillFlow\-Benchexposes this qualitatively through incorrect\-skill drift and skill inflation, andSkillForgedoes so in a deployment\-style failure\-diagnosis loop, but the field still lacks a standard scalar repair metric\.

### 8\.3Four Comparisons That Remain Hard

Benchmark progress does not yet make the literature head\-to\-head comparable\.*Cross\-operator comparison*lacks operator velocities\.*Cross\-library\-size comparison*lacks smooth size sweeps\.*Cross\-backbone comparison*is confounded by model, harness, context length, and tool surface, even in useful early transfer studies such asCoEvoSkillsandXSkill\.*Cross\-task\-distribution comparison*remains immature: adjacent protocols such asELL/StuLife\(Caiet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib71)\),LifelongAgentBench\(Zhenget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib70)\), andProEvolve\(Liet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib66)\)expose lifelong, proactive, or programmable shift, but most deployment papers still use bespoke task streams\. These limits determine how the next section treats evidence: the patterns in §[9](https://arxiv.org/html/2607.10113#S9)rely mainly on within\-paper ablations and convergent benchmark behavior rather than cross\-paper leaderboards\.

### 8\.4Toward Trajectory\-Aware Evaluation

A trajectory\-aware protocol should report the library as a time series\. Four elements are sufficient: performance, skill count, and retrieval quality at a grid of task indices or library sizes; operator velocities forAdd,Refine,Merge, andPrune; a drift schedule or family\-transfer condition; and a maintenance\-off or repair\-off ablation\.Single\-Agent\-Skills’ size sweep,SkillRouter’s 80K routing benchmark,SRA’s decomposition of retrieval versus incorporation,SWE\-Skills\-Bench’s with\-skill/without\-skill deltas,SkillFlow\-Bench’s skill\-count trajectories,SkillsVote’s skill\-linked credit assignment, andRaw\-Experience’s extractor\-consumer split are complementary starting points\.

SkillFlow\-Benchmoves the field in this direction, but its family\-reset design leaves open how one global library behaves under heterogeneous workflows\. The next step is to add cross\-family global\-library protocols, operator\-velocity logging, and maintenance\-off ablations\.

## 9Seven Evidence\-Graded Patterns

Building on the evaluation audit in §[8](https://arxiv.org/html/2607.10113#S8), the primary dynamic\-skill cluster within the 124\-paper modern audit set surfaces seven recurring patterns across methods, benchmarks, and artifact types\. They should not be read as pooled effects\. Evidence is strongest for within\-paper ablations, moderate for convergent benchmark behavior, and weakest for architectural corroboration without ablation\.

Table[7](https://arxiv.org/html/2607.10113#S9.T7)summarizes the evidence for the section and assigns each pattern the evidence grade defined in §[2\.5](https://arxiv.org/html/2607.10113#S2.SS5); §[12](https://arxiv.org/html/2607.10113#S12)uses these patterns to frame the open\-problem agenda\.

Table 7:Seven evidence\-graded patterns in the dynamic\-skills literature \(2023–2026\)\.Each row summarizes an observed pattern \(§[9](https://arxiv.org/html/2607.10113#S9)\), an evidence grade using the convention in §[2\.5](https://arxiv.org/html/2607.10113#S2.SS5), the methods that provide supporting ablations or measurements, and the main evidence boundary\. Grades are deliberately conservative: they reflect heterogeneous backbones, task surfaces, evaluators, and artifact types rather than pooled effect sizes\. The strongest themes concern write\-time discipline; the retrieval\-scaling and focus rows argue for smaller libraries from the retrieval\-resolution and distractor\-load ends respectively; the weaker\-backbone row is suggestive and deployment\-facing\. The open\-problem agenda in §[12](https://arxiv.org/html/2607.10113#S12)is organized around these patterns\.If the analysis is restricted to the primary dynamic\-skill systems and excludes memory\-only, registry\-only, and purely parametric boundary cases, the strongest patterns are admission, verifier quality, and maintenance/repair\. Retrieval scaling remains moderately supported because it has both a controlled size sweep and registry\-scale corroboration, but it still lacks a shared benchmark across storage designs\. Weaker\-backbone gains, focused\-library advantage, and write\-time abstraction are useful cross\-system signals rather than settled effects; they are retained because they describe recurring design pressure, not because they support a pooled effect size\.

### 9\.1Curated skills outperform unverified self\-generated skills

One of the clearest within\-paper findings in the surveyed corpus is that*admission matters*\. Libraries whose write\-time verifier is stronger than “the agent proposed it” usually beat libraries that admit any proposal, and the evidence spans methods at both ends of the verification spectrum\.EvoSkill\(Alzubiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib15)\)reports that its Pareto\-front selection over held\-out validation performance avoids the redundant/conflicting\-skill accumulation seen under greedy acceptance and yields\+7\.3\+7\.3absolute points on OfficeQA \(Table 1 / Figure 2\) and\+12\.1\+12\.1points on SealQA;CoEvoSkills\(Zhanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib16)\)provides the cleanest verifier ablation in the corpus, with Table B1 showing that removal of the surrogate verifier drops SkillsBench pass rate from71\.1%71\.1\\%to41\.1%41\.1\\%;SkillWeaver\(Zhenget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib13)\)reports that removing the “practice \+ verify” phases of PPVH \(leaving only propose\+hone\) drops benchmark performance below the no\-skill\-library baseline on its hardest web tasks; andTrace2Skill\(Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22)\)reports that prevalence\-weighted consolidation is useful only when consolidation also filters on a judge score\. New executable web and GUI systems strengthen the same point:ASI\(Wanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib100)\)gains over text skills partly because induced programs are execution\-verified before becoming actions,WebXSkill\(Wanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib101)\)reports that validation and curation are necessary for its grounded/guided web\-skill gains, andContractSkill\(Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102)\)improves self\-generated web skills by turning failure into verifier\-localized patch admission\. Late\-May optimization papers sharpen the same mechanism:SkillGenselects candidate skills by net interventional effect,SkillOptaccepts text edits only when held\-out validation improves, andSkillMastertrains skill edits with counterfactual utility rewards\(Maet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib119); Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128);[a](https://arxiv.org/html/2607.10113#bib.bib117)\)\. Additional 2026 systems strengthen the same point from new domains:SkillFoundry\(Shenet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib74)\)only admits scientific packages after contract/provenance/test validation,SkillsVote\(Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125)\)admits only successful reusable discoveries after skill\-linked credit assignment, andSkillForge\(Liuet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib75)\)improves deployed support skills through failure analysis, diagnosis, and optimization rather than blind rewriting\. The pattern also appears in maintenance\-heavy systems:AutoRefine\(Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25)\)’s pruning/merging gate andAgentSkillOS\(Liet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib42)\)’s audit\-gated organization both report that disabling the gate erodes downstream task success\.

#### Caveats\.

This pattern says that replacing no verification with some verification is usually valuable, not that more verification is always better\. The main exception is the memory/trajectory family \(SimpleMem, parts ofMUSE\)\(Yanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib9)\), where artifacts are episodic cases and admission is closer to relevance filtering than quality control\.

### 9\.2Verification*quality*is often decisive in skill\-aware RL

Inside skill\-aware RL, the*quality of the verifier or reward\-shaping signal*is often one of the most load\-bearing choices\.CoEvoSkills’ Table B1 surrogate\-verifier ablation isolates a3030\-point verifier effect;CODE\-SHARPimproves success from24\.30%24\.30\\%to41\.02%41\.02\\%through refinement mutations under a learned gate;Agentic\-Proposing’s verifier ensemble plus dynamic pruning improves problem validity from68\.7%68\.7\\%to82\.3%82\.3\\%;Co\-Evolving\(Junget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib36)\)attributes substantial gain to hard\-negative construction; andSELAUR\(Zhanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib30)\)shows that uncertainty\-aware reward shaping can matter as much as the RL algorithm\.SkillMasteradds the clearest new variant: skill edits receive counterfactual probe utility and separate advantage normalization from task actions\(Yanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib117)\)\.SkillFlow\-Recursive\(Zhanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib121)\)is architectural corroboration from flow matching, where recursive skill evolution uses trajectory\-balance diagnostics rather than direct prompt judgment\.ASG\-SI\(Huang and Huang,[2025](https://arxiv.org/html/2607.10113#bib.bib55)\)specifies the audited graph and verifiable\-reward machinery a deployed system would need\.

#### Caveats\.

This pattern is specific to skill\-aware RL\. In executable libraries, once execution verification exists, additional rollouts or tighter contracts may have diminishing returns relative toPruneandRerank\.

### 9\.3Flat retrieval often degrades at moderate library sizes

Flat skill libraries can show an accuracy drop at moderate sizes, consistent with a top\-kkretrieval signal\-to\-noise collapse\.Single\-Agent\-Skills\(Li,[2026](https://arxiv.org/html/2607.10113#bib.bib52)\)gives the clearest controlled size sweep: selection remains high at 16–64 skills, degrades around 128 skills, and drops sharply at 256 skills unless hierarchical routing or disambiguating instructions are added\.AgentSkillOS\(Liet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib42)\)motivates its capability tree with flat\-invocation collapse,Wild\-Skills\(Liuet al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib27)\)shows the same mechanism under forced loading, autonomous selection, distractors, 34K\-pool retrieval, and no curated task\-specific skills, andSkillRouter\(Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43)\)shows that 80K\-scale registries need full\-text retrieve\-and\-rerank because metadata\-only routing loses implementation\-level signals\.SRA\(Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)adds a decomposed 26K\-corpus benchmark: better retrieval helps, but incorporation and need\-aware loading remain separate bottlenecks\.Graph of Skills\(Liuet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib89)\)adds 200–2,000\-skill evidence that dependency\-aware graph retrieval can preserve reward while compressing token use\.SkillsInjectorgives the read\-time analogue: static all\-injection can collapse as candidate pools grow, while adaptive budgeting and set\-aware rendering improve downstream pass rates\(Liet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib131)\)\.

library sizeselection accuracy \(%\)9778641632641282569796927864hierarchy / graph / rerankκ≈50\\kappa\\\!\\approx\\\!50–100100controlled size sweepWild\-Skills: 34K\-pooldistractor stressSkillRouter: 80Kfull\-text rerankGraph of Skills:200–2,000 skills

Figure 4:Retrieval\-scaling evidence behind R3\.The blue curve plots the controlledSingle\-Agent\-Skillssweep reported in the paper: 16–32 skills remain around9696–98%98\\%, 64 skills at92%92\\%, 128 skills at78%78\\%, and 256 skills at64%64\\%\. The dashed green path summarizes the mitigation family rather than a pooled leaderboard: hierarchy, graph retrieval, and full\-text reranking push the failure rightward inSingle\-Agent\-Skills,Wild\-Skills,SkillRouter, andGraph of Skills, but they are not yet comparable under one shared harness\.#### Caveats\.

The location of the drop depends on retriever quality, description length, skill orthogonality, and whether the corpus exposes full skill bodies\. Hierarchical, DAG, and ontology storage can push the drop rightward \(AgentSkillOS,SkillX,SkillOrchestra\)\(Wanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib46)\); full\-text reranking \(SkillRouter\) addresses the complementary registry\-scale routing problem\(Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43)\)\. The literature has not shown general removal\.

### 9\.4Several studies report larger relative gains for weaker backbones

Several studies report larger relative gains for weaker base models than for stronger ones, but the evidence is convergent rather than a controlled cross\-paper effect\.SkillWeaver\(Zhenget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib13)\)closes a larger fraction of the capability gap on weaker web\-agent backbones;MetaClaw\(Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35)\)reports larger relative two\-timescale gains on weaker backbones in its setting;EvoSkillreports analogous gap compression; andAgentic\-Proposing’s headline result is on a 30B model, with a smaller relative lift on the frontier teacher\. Because these papers vary in harness, context budget, and gain definition, we treat the pattern as a deployment\-facing hypothesis rather than as a comparable effect\-size estimate\.

#### Caveats\.

The effect depends on library content\. Common\-but\-not\-obvious heuristics help weaker models more; rare specializations can invert the pattern if weaker models cannot route to them reliably\.XSkill’s Qwen transfer results are the caution: transfer can raise Pass@44by encouraging exploration while lowering Average@44\.

### 9\.5Focused libraries often beat comprehensive ones

A library curated to a narrow task distribution can outperform a more comprehensive library, even when the former is a subset of the latter\. This is the content\-level dual of retrieval degradation: the failure mechanism is distractor load\.SkillX\(Wanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib21)\)reports that task\-family\-tuned hierarchies beat flat libraries;Wild\-Skillsshows that retrieval\-utility filtering beats skills that looked useful in isolation; andCASCADE\(Huanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib20)\)reports that consolidation beats pure growth at fixed compute\.SkillMOO\(Gonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib76)\)optimizes pass rate and cost jointly,SkillLearnBench\(Zhonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib82)\)shows that plausible skill text does not imply competence,SWE\-Skills\-Bench\(Hanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib104)\)reports only\+1\.2%\+1\.2\\%average gain across public SWE skills with many zero\-utility or negative cases, andSkillFlow\-Bench\(Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48)\)finds that stronger settings maintain compact revised skills while weaker settings often fragment into many low\-utility files\.SkillsInjectorandRaw\-Experienceadd the context and extraction views: a larger injected set can hurt, and a skill useful for one consumer can transfer negatively to another\(Liet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib131); Huanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib127)\)\. The single\-skill analogue is the SKILL\.md focused\-prompt result: trimming L2 prose to task\-relevant instructions improves use\.

#### Caveats\.

In non\-stationary deployments \(SkillClaw,AutoSkill,MetaClaw\), focused libraries can become stale; the result then argues for fastPrune, not permanent narrowness\.

### 9\.6At moderate\-to\-large library sizes, maintenance becomes load\-bearing

Systems with at least one negative or repair operator—Prune,Merge,Refine, or an equivalent—often beat monotonic\-growth systems once distractor load matters\.AutoRefine\(Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25)\)reports the headline ALFWorld result in its Table 1, but its maintenance\-off evidence is the TravelPlanner validation ablation: removing periodic pruning and merging lowers final pass rate from35\.6%35\.6\\%to31\.1%31\.1\\%, grows the repository by4\.5×4\.5\\times, and reduces utilization from0\.710\.71to0\.080\.08\(its Figures 2–3\)\.SkillOps\(Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\)gives a direct library\-time variant: removing library\-time maintenance drops standalone ALFWorld success to71\.9%71\.9\\%, while the full system reports79\.5%79\.5\\%and plug\-in gains of\+0\.68\+0\.68–\+2\.90\+2\.90points for retrieval\-heavy baselines\.ContractSkillshows the single\-artifact analogue: local verifier\-guided patch repair beats naive self\-generated web skills on VisualWebArena\(Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102)\)\.SkillDroidshows the repeated\-use analogue: failure\-count recompilation lets a mobile GUI skill library improve while a stateless baseline degrades under instruction variation\(Chenet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib103)\)\.EmbodiSkilladds an embodied repair caveat: the system must distinguish skill defects from execution lapses before rewriting valid guidance\(Juet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib118)\)\.AgentSkillOSattributes scaling to periodic capability\-tree audit;Wild\-SkillsandEffiSkill\(Wanget al\.,[2026i](https://arxiv.org/html/2607.10113#bib.bib53)\)find utility\-based filtering can beat increased admission; andPSN’s maturity gating loses its stated advantage when disabled\.SkillFlow\-Benchadds that weaker settings often fail by skill inflation and incorrect\-skill drift rather than inability to write skills\.SkillGradandSkillOptframe repair as iterative text\-space optimization rather than one\-shot reflection\(Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130); Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128)\)\. Parametric systems express the same point through distillation windows \(MetaClaw,SKILL0\)\(Luet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib31)\): mistimed consolidation lets the fast\-loop library grow without bound\.

#### Caveats\.

The exception is very short horizons, where maintenance cost has no time to amortize\.

### 9\.7Write\-time abstraction is usually the stronger backbone than read\-time abstraction alone

The literature usually favors*write\-time*abstraction—retrospective induction,Abstract, orDistillduring skill authoring—over read\-time summarization alone\.Trace2Skill\(Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22)\)’s prevalence\-weighted consolidation beats a read\-time summarizer baseline;CASCADE’s write\-time distillation beats a retrieval\-plus\-rerank control; andSimpleMem\(Liuet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib58)\)shows in Table 5 that removing write\-time semantic compression drops LoCoMo average F1 from43\.2443\.24to31\.2931\.29\.XSkill\(Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12)\)extends this finding to multimodal agents: its Table 3 ablation shows that removing the Experience Manager drops VisualToolBench Average@44by4\.094\.09points and removing the Skill Manager drops it by3\.623\.62, while read\-time task\-decomposition/adaptation ablations are smaller\.SkillFlow\-Benchsupports file\-level write\-time abstraction but is not a clean write\-time\-versus\-read\-time ablation;SkillGen,SkillOpt, andSkillGradadd direct evidence that skill quality improves when success/failure evidence is converted into a persistent artifact before deployment; andRaw\-Experienceshows that extraction quality and consumption quality must be evaluated separately\(Maet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib119); Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128); Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130); Huanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib127)\)\.K2\-Agent\(Wuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib34)\)’s declarative\-procedural split is architectural corroboration\.

#### Caveats\.

The result is cleanest under reasonably stationary tasks\. In non\-stationary deployments and visually grounded multimodal settings, read\-time adaptation layered on top of write\-time abstraction can still be necessary\.SkillTTA\(Wanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib123)\)is the boundary case: it synthesizes task\-specific transient skills at read time, improving several benchmarks without maintaining a persistent library\.

### 9\.8Cross\-pattern observations

Three observations connect the patterns\. Admission, verifier quality under RL, maintenance, and write\-time abstraction are all forms of*write\-time discipline*\. Retrieval scaling and focused\-library effects both argue that smaller can be better, through different mechanisms: retrieval resolution and distractor load\. Finally, lifecycle benchmarks warn that skill*usage*is not skill*utility*;Wild\-Skillsshows this through retrieval realism, andSkillFlow\-Benchthrough high\-use, low\-gain settings\. The open problems in §[12](https://arxiv.org/html/2607.10113#S12)either exploit these patterns or ask why a method appears to violate one without visible cost\.

## 10Infrastructure

Dynamic\-skill methods depend on packaging formats, storage backends, marketplaces, SDKs, and edit\-execution pipelines\. These choices were peripheral in 2023 but are load\-bearing by 2026: they define the skill unit, determine scaling limits, shape cross\-user aggregation, and decide which operators can be implemented at useful velocity\.

Table[8](https://arxiv.org/html/2607.10113#S10.T8)groups representative systems by structural constraints, not quality\. The visible diagonals—flat storage with near\-monotonic\{Add\}\\\{\\textsc\{Add\}\\\}, hierarchies withSplit\-heavy maintenance, and DAG/ontology stores withCompose\-heavy workflows—explain why methods rarely transfer cleanly across infrastructure stacks: off\-diagonal operators are often too expensive to run at the required edit velocity\.

Dim\.ClassRepresentative systemsFast opsaCostly opsaPrimary trade\-offPackageSKILL\.mdSkillClaw,AutoSkill,MetaClaw,MUSE\-Autoskill,Skilldex, public Agent Skills specAdd,Refine\\textsc\{Add\},\\textsc\{Refine\}Composeportability/lock\-inTool schemaFunction\-calling \+ MCP/plugin manifestsAdd,Compose\\textsc\{Add\},\\textsc\{Compose\}Abstractschema/capabilityHybrid skill/memoryXSkill,WebXSkill,Memento,SAGERbAdd,Rerank\\textsc\{Add\},\\textsc\{Rerank\}Merge,Prune\\textsc\{Merge\},\\textsc\{Prune\}portability/compat\.Compiled corpusCorpus2Skill,SkillRepoMining,SkillFoundry,SkillSmithAdd,Abstract\\textsc\{Add\},\\textsc\{Abstract\}Pruneprovenance/qualityTrajectorySimpleMem,MUSE,SkillDroidbAdd,Rerank\\textsc\{Add\},\\textsc\{Rerank\}Abstract,Distill\\textsc\{Abstract\},\\textsc\{Distill\}read\-time/write\-timeStorageFlat embeddingEvoSkill,SkillWeaver,Agentic\-ProposingAddMergeadmit/reorganiseHierarchical treeAgentSkillOS,SkillX,Corpus2Skill,Uni\-SkillSplitPrunerestructure/removeGraph / dependencyPSN,CODE\-SHARP,Graph of Skills,GraSP,WebXSkill,SSL,SkillOpsComposecPrunechain/removeTyped ontologySkillOrchestraCompose\(typed\)Refineexpressivity/costMarketPull\-modelCentral curated registry;Skilldex,SkillsVoteAdd,Rerank\\textsc\{Add\},\\textsc\{Rerank\}Mergecuration/scalePush\-modelSkillClawcross\-userAddMerge,Prune\\textsc\{Merge\},\\textsc\{Prune\}velocity/provenanceHybridAutoSkilldual reviewAdd,Rerank\\textsc\{Add\},\\textsc\{Rerank\}Merge,Prune\\textsc\{Merge\},\\textsc\{Prune\}review/throughputSecurity scannerSkillSieve,Malicious\-or\-Not,Semantic\-Supply,Credential LeakagePruneRefinerecall/false positivesPipelineInline editPSN,CODE\-SHARPRefine,Rewrite\\textsc\{Refine\},\\textsc\{Rewrite\}Prunevelocity/rollbackLog\-and\-applyAutoRefine,AgentSkillOS,SkillGrad,SkillOpsPruneRefineaudit/latencyCompile\-and\-serveSkillSmithAbstractRefinecompactness/fidelityShadow exec\.MetaClaw,SKILL0DistillRefinesafety/stalenessRelease auditMedSkillAudit,AgentDevelPruneRefinerigor/throughput

- a“Fast” / “costly” are relative within a class: listed entries are drawn from the ten\-operator vocabulary \(§[7\.1](https://arxiv.org/html/2607.10113#S7.SS1)\) and flag which operators a given infrastructure stack admits cheaply \(unit\-cost edits\) versus which require a heavier mechanism \(locking, re\-indexing, cross\-user consensus, or human review\)\.
- bSimpleMemandMUSEare trajectory\-memory systems in which episodes themselves play the skill\-like role of retrievable artifacts;SkillDroidcompiles trajectories into executable GUI templates;XSkillandMementoare hybrid cases that pair skill documents with case / experience stores\. We include these methods when their artifact is read by an agent loop as if it were a skill\. The central dynamic\-skills\-vs\-memory distinction is preserved by the*Artif\.*and*Trig\.*columns of TableLABEL:tab:master\.
- cForPSNandCODE\-SHARP, the graph is cheap for library\-edit\-levelComposeto*audit*\(prerequisites are explicit\); admitting a composed skill as a new library entry still requires a verifier pass, so the cheapness is structural, not free\.

Table 8:Four infrastructure dimensions \(package, storage, market, pipeline\) that materially constrain dynamic skill systems\.Each row classifies a representative deployment along the same categorical vocabulary so that the operator economy and the primary trade\-off are readable at a glance\. The “fast ops” / “costly ops” asymmetry is the core point of the table: infrastructure choices predetermine which operator vocabulary is economical, which is why the operator–storage diagonal of the master taxonomy \(TableLABEL:tab:master\) is sparsely populated\.### 10\.1Packaging formats and the minimum viable unit

The surveyed literature uses four packaging formats\.SKILL\.mdpackages dominate the agent\-skill regime: one directory per skill, with a descriptor, optional scripts, and an interface manifest\. Public product/specification sources define the practical convention as a folder withSKILL\.mdmetadata and progressive disclosure from name/description to full instructions and resources\(Zhanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib115); Anthropic,[2025](https://arxiv.org/html/2607.10113#bib.bib116)\)\. Function\-calling schemas and MCP/plugin manifests dominate tool\-as\-skill systems, where the portable object is a typed callable signature\. Hybrid skill\-memory stores, as inXSkill,WebXSkill, andMUSE\-Autoskill\(Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12); Wanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib101); Linet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib129)\), pair a workflow with executable or experience evidence;MUSE\-Autoskillmakes this explicit through per\-skill memory\. A fourth, less explicit format treats trajectories themselves as retrievable artifacts \(SimpleMem,MUSE\) or compiles them into GUI templates \(SkillDroid\)\(Chenet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib103)\)\. Recent infrastructure papers add package\-manager, corpus\-compiler, compiler\-runtime, typed\-contract, and structured\-representation variants:Skilldex\(Saha and Hemanth,[2026](https://arxiv.org/html/2607.10113#bib.bib81)\)gives skills scoped package semantics,Corpus2Skill\(Sunet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib79)\)compiles an enterprise corpus into a navigable skill directory,SkillSmith\(Xuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib122)\)compiles skill packages into boundary\-guided runtime interfaces,SkillOps\(Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\)casts each executable skill as a contract with preconditions, operation, artifacts, validators, and failure modes, andSSL\(Lianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib112)\)normalizes skill text into scheduling, structural, and logical fields for retrieval and risk review\.

The formats trade portability against structure\.SKILL\.mdtravels across backbones but depends on the receiver’s ability to follow prose; function schemas travel across models but assume the target tool exists; hybrid stores require both workflow and experience interfaces; trajectory memories grow freely but are task\-specific\.K2\-Agent’s declarative\-procedural pair andXSkill’s Markdown\-plus\-experience split are the closest attempts to bridge formats, but both still assume compatible tools and context interfaces\.

### 10\.2Storage backends and scaling limits

Storage determines whether moderate\-library\-size retrieval degradation is conceded or delayed\. Flat embedding indices \(EvoSkill,SkillWeaver,Agentic\-Proposing\) use dense retrieval over descriptions and are the class most directly implicated by the scaling evidence\. Hierarchical stores \(AgentSkillOS,SkillX,Corpus2Skill,Uni\-Skill\) push the degradation point outward; graph stores \(CODE\-SHARP,PSN,Graph of Skills,GraSP,WebXSkill,SSL,SkillOps\) expose skill–skill, page–skill, or representation\-level relations to retrievers and audits; ontology stores \(SkillOrchestra\) attach semantic categories for typed composition\.SkillOpsadds a maintenance\-specific graph layer in which dependency, compatibility, redundancy, and alternative edges drive health diagnosis before downstream retrieval\.SRAadds the benchmark view: a large skill corpus is not only a retrieval target but also an incorporation problem, because a correct top\-kkset does not ensure need\-aware loading\(Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)\.

For dynamic systems, the key question is operator cost under the storage class\. Flat indices makeAddcheap andMergeexpensive; hierarchies makeSplitcheap and subtreePruneexpensive; DAGs makeComposeeasier to audit butPrunerisky because descendants may depend on the removed node; ontologies make category inheritance cheap but category\-changingRefineexpensive\. This operator–storage coupling explains why the master table’s off\-diagonal combinations remain sparse\.

### 10\.3Marketplaces, plugins, and cross\-user aggregation

A 2026 development is the skill*marketplace*: a registry where skills authored by one user or team are listed, reviewed, ranked, and adopted by others\. Marketplaces change admission from a single verifier score to an aggregate utility and moderation signal\.Agent Skills: Data\-Driven Analysis\(Linget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib87)\)measures the speed of the public Claude\-style skill ecosystem;SkillsVote\(Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125)\)adds governed collection, recommendation, skill\-linked credit, and evidence\-gated evolution; andAgentSkills\-Wild\(Liuet al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib106)\),Malicious\-Skills\-Wild\(Liuet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib107)\),Malicious\-or\-Not\(Holzbaueret al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib95)\),Semantic\-Supply\(Sahaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib120)\), andCredential Leakage\(Chenet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib92)\)show why registry metadata, repository context, behavioral verification, semantic retrieval manipulation, and remediation workflows are now infrastructure, not a separate security appendix\.

The literature distinguishes pull\-model registries with central admission, push\-model sharing as inSkillClaw, hybrid registries with per\-user overrides as inAutoSkill, governed evolution as inSkillsVote, and package\-manager models such asSkilldexwith scoped installs and conformance checks\. Pull models foreground retrieval at scale; push, hybrid, governed, and package\-manager models foreground attribution, credit assignment, and governance\.

### 10\.4The edit–execute pipeline

Each operator in §[7\.1](https://arxiv.org/html/2607.10113#S7.SS1)must be implemented as an edit to a concrete artifact\. Three pipeline architectures recur\.*Inline editing*\(PSN,CODE\-SHARP,ContractSkill\) edits the artifact in place and needs an explicit version store for rollback\.*Log\-and\-apply*pipelines \(AutoRefine,AgentSkillOS,SkillDroid,SkillGrad,SkillOps\) stage or schedule the edit, verify it, and commit only after admission, successful replay, or a health\-triggered maintenance pass\.*Shadow\-execution*pipelines \(MetaClaw,SKILL0\) evolve a shadow library while serving from the main one, then cut over during a distillation window\. A fourth,*compile\-and\-serve*pattern is emerging inSkillSmith: a larger skill package is compiled into a smaller boundary\-guided runtime interface before repeated use\.

These pipelines determine whether a proposed operator is cheap enough to use continuously or expensive enough to reserve for release gates\.

## 11Safety and Governance

The 2026 safety wave turns dynamic skills from a prompt\-injection concern into a software\-supply\-chain problem\. The corpus now includes attack taxonomies, registry\-scale measurements, admission scanners, confirmed malicious\-skill datasets, credential\-leakage studies, harmful\-skill benchmarks, request\-conditioned invocation audits, repository\-context audits, and skill\-stealing attacks\. Because skill artifacts combine natural language, code, configuration, sometimes learned models, and social provenance, admission control, provenance, sandboxing, and release governance must be lifecycle stages\.

Table[9](https://arxiv.org/html/2607.10113#S11.T9)summarizes the evidence\. The field has attack studies and partial defenses, but few defended lifecycle systems: methods verify utility but not maliciousness, scanners analyze artifacts but not downstream composition, and registries expose popularity but not operator\-level lineage\. Dynamic skill stores need a security pipeline that mirrors their lifecycle pipeline\.

Table 9:Safety surfaces for dynamic skill systems after the 2026 safety wave\.The table distinguishes attack evidence, defense evidence, and governance primitives\. Skill safety is not a single prompt\-injection benchmark: it includes supply\-chain poisoning, harmful\-but\-valid capability packages, credential leakage, scanner calibration, invocation\-time risk, skill theft, and domain release readiness\.### 11\.1Eight safety surfaces specific to dynamic skills

The surfaces below are specific to, or strongly amplified by, lifecycle\-managed skill stores\.Towards Secure Agent Skills\(Liet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib72)\)organizes threats across creation, distribution, deployment, and execution; we map that phase structure onto the survey’s operators and artifact\-store framing\.

#### Prompt injection through admitted skills\.

An admitted skill can carry adversarial instructions in documentation, examples, comments, or generated helper files\.AgentSkills\-PI\(Schmotzet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib105)\)establishes the basic attack channel for Claude\-style skill folders: skill metadata routes the agent to a file whose body or referenced scripts may contain hidden instructions\.ClawSafety\(Weiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib56)\)quantifies the outcome\-level risk: malicious skill files are the highest\-ASR injection vector in its OpenClaw personal\-agent benchmark, averaging69\.4%69\.4\\%ASR across backbones\.Skill\-Inject\(Schmotzet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib108)\)adds benchmark\-scale evidence with 202 injection\-task pairs and contextual ASR reaching roughly the 40–80% range depending on model and setting\.SkillJect\(Jiaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib109)\)shows that attackers can optimize the skill artifact itself, reporting 95\.1% average ASR after trace\-driven closed\-loop refinement versus 10\.9% for naive injections\.SkillAttack\(Duanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib78)\)shows the complementary red\-team view: benign\-looking skills can become exploitable when an attacker refines prompts into attack paths, with injected adversarial skills reaching high ASR and real\-world Hot100 skills still exploitable under some conditions\.

#### Supply\-chain poisoning at publication and installation\.

Skill packages can be poisoned before a user ever invokes them\.AgentSkills\-Wild\(Liuet al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib106)\)measures broad vulnerability prevalence in public registries, analyzing 31,132 skills and reporting that 26\.1% contain at least one vulnerability pattern; executable\-script skills are 2\.12×\\timesmore likely to contain vulnerabilities than instruction\-only skills\.Malicious\-Skills\-Wild\(Liuet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib107)\)provides confirmed\-malice evidence: from 98,380 public skills, the authors identify 157 behaviorally confirmed malicious skills with 632 labeled vulnerabilities and report 93\.6% removal after responsible disclosure\.Supply\-Chain\-Poisoning\(Quet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib73)\)constructs DDIPE attacks over 1,070 adversarial skills across 15 MITRE ATT&CK categories and reports bypass rates from 11\.6% to 33\.5% depending on defense configuration, with 2\.5% evading both static and alignment filters\.Semantic\-Supply\(Sahaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib120)\)adds the registry\-facing semantic channel: SKILL\.md triggers and descriptions manipulate discovery, selection, and governance even when no executable exploit has run\.BadSkill\(Tieet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib91)\)adds a distinct model\-in\-skill threat: a skill can bundle a backdoored model whose malicious behavior is hidden in learned parameters, reaching 97\.5–99\.5% ASR while largely preserving benign accuracy\.

#### Harmful but well\-formed skills\.

A skill can be valid, useful, and non\-poisoned while still enabling harmful functionality\.HarmfulSkillBench\(Jianget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib80)\)finds 4,858 harmful skills among 98,440 collected skills and shows that installed harmful skills can raise harm scores even under implicit invocation\. This is a policy\-governance problem rather than malware detection\.

#### Credential and secret leakage\.

Credential leakage is a cross\-modal property of skills, because secrets can be exposed only through the interaction of SKILL\.md text, code, runtime stdout, and agent context\.Credential Leakage\(Chenet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib92)\)samples 17,022 skills from a 170,226\-skill SkillsMP population, identifies 520 affected skills and 1,708 issues, and reports that 76\.3% of cases require joint NL\+code analysis\. Debug logging accounts for 73\.5% of vulnerability issues because stdout is often fed back to the LLM, turning routine logs into credential disclosure\.

#### Admission scanner limits and false positives\.

Security scanners are necessary but not sufficient\.AgentSkills\-Wildreports a broad SkillScan detector with 86\.7% precision and 82\.5% recall, whileSkillSieve\(Hou and Yang,[2026](https://arxiv.org/html/2607.10113#bib.bib77)\)is one of the strongest current defense\-side triage results, filtering 86% of skills at zero API cost and reporting 0\.800 F1 at about $0\.006 per skill\.Malicious\-or\-Not\(Holzbaueret al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib95)\)shows the opposite failure mode: scanner\-only classification can badly overstate risk, while repository context reduces 2,887 scanner\-flagged skill\-repository combinations to 0\.52% remaining in malicious flagged repositories\.STARS\(Zhanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib110)\)adds the invocation\-time view: the risk of calling a skill depends on the request and context, and its calibrated fusion model improves held\-out indirect\-prompt\-injection HR\-AUPRC over static\-only and contextual\-only scorers\.User\-Comprehension\(Wen,[2026](https://arxiv.org/html/2607.10113#bib.bib126)\)adds a human\-facing audit gap: many skill specifications do not expose enough operational basis, output contract, boundary disclosure, or examples for users to form bounded expectations\. Admission gates therefore need content analysis, provenance/context analysis, request\-conditioned invocation review, and user\-facing capability disclosure\.

#### Ownership, confidentiality, and skill theft\.

Once skills become valuable artifacts, the market creates confidentiality and intellectual\-property risks\.SkillStealing\(Wanget al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib85)\)studies black\-box extraction from proprietary agents and reports that paid or proprietary skill behavior can be inferred with only a few interactions, with substantial semantic leakage even when exact text recovery is low\. This surface is orthogonal to user harm: the victim may be a skill author or registry operator rather than the end user\.

#### Domain\-specific release readiness\.

High\-stakes domains need release gates beyond generic maliciousness\.MedSkillAudit\(Houet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib93)\)evaluates 75 medical research skills, finds that 57\.3% fall below the Limited Release threshold, and reports automated\-audit ICC\(2,1\)=0\.449 against expert consensus\. The result supports domain\-specific rubrics for scientific integrity, reproducibility, and boundary safety\.

#### Attribution loss and composition misuse\.

Maintenance operators can delete, merge, split, and rewrite skills\. Without operator\-level provenance, later audits cannot reconstruct which artifact contributed to a decision\. Composition adds another risk: individually acceptable skills can be harmful when chained\.ASG\-SI\(Huang and Huang,[2025](https://arxiv.org/html/2607.10113#bib.bib55)\)is the closest surveyed system to graph\-level audit, andSkillOrchestra\(Wanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib46)\)gestures toward typed composition, but neither closes the full provenance\-plus\-composition safety loop\.

### 11\.2Coverage gap across the surveyed literature

The table’s main implication is that skill safety is a lifecycle property: the vulnerable point can be admission, publication, installation, execution, marketplace review, or later composition\.

### 11\.3Governance primitives a deployment would need

A defensible deployment needs eight primitives: an*admission\-time scanner*combining static checks, LLM/rubric analysis, sandboxing, and repository provenance; a*semantic registry monitor*for trigger/description manipulation in discovery and selection; a*package manifest*for dependencies, permissions, model artifacts, network endpoints, and allowed tools; a*request\-conditioned invocation audit*for deciding whether a skill should be called in the current context; a*runtime containment layer*for stdout, filesystem, network, and bundled\-model channels; an*operator\-granular provenance log*forAdd,Refine,Merge,Split, andPrune; a*policy and domain release gate*; and a*market governance layer*for ownership, takedown, remediation, abandoned\-repository hijacking, and user\-facing capability disclosure\.

If a system can create, modify, distribute, retrieve, and compose skills automatically, it must also audit those operations automatically\.

## 12Open Problems

The seven evidence\-graded patterns constrain the next research agenda: new methods should exploit them or explain why their setting violates them\. This section states eight open problems, organized around admission, maintenance, retrieval/composition, and deployment\. We label each problem’s evidence basis as*measured*when it follows from direct benchmark or deployment evidence,*extrapolated*when it extends measured behavior beyond the tested regime, and*speculative*when the literature has not yet run the necessary horizon or scale\. Table[10](https://arxiv.org/html/2607.10113#S12.T10)gives a concrete first experiment for each\.

Table 10:Eight open problems for the dynamic\-skills research community, each paired with a concrete first experiment\.The first four rows are method\-level problems tractable for a single team; the last four require infrastructure or community coordination\. The “likely obstacle” column states the main technical blocker for planning a research program\.### 12\.1Compositional verifiers for code skills

*Evidence basis: extrapolated from measured verifier ablations\.*Verifier design is decisive in skill\-aware RL, andPSN’s rollback\-gated behavioral validation andCODE\-SHARP’s learned judges expose a useful tradeoff: probe\-set behavioral coverage versus broader but less grounded judge coverage\. No surveyed method composes them\.

### 12\.2Admission under non\-stationary task distributions

*Evidence basis: extrapolated from deployment and focus/maintenance evidence\.*Admission and focus interact badly under non\-stationary deployment: yesterday’s verifier can become a poor predictor of today’s utility, and yesterday’s focused library can become harmful\. Existing systems \(SkillClaw,AutoSkill,MetaClaw\) use eviction but do not adapt the verifier itself\. The open problem is a verifier whose threshold and scoring function update from downstream utility\.

### 12\.3Principled maintenance schedules

*Evidence basis: measured need, extrapolated scheduling model\.*Maintenance becomes important once libraries grow, but published schedules are engineering defaults\.AutoRefineruns per\-episode maintenance,MetaClaw/SKILL0/K2\-Agenttrigger distillation periodically or by prevalence, andAgentSkillOSaudits on a human\-scale cadence\. The open problem is to derive when to invokePrune,Merge, orDistillfrom monitored quantities such as admission rate, drift rate, retrieval entropy, and maintenance cost\. Online learning with restarts is one possible analogue\.

### 12\.4Preventing parametric\-collapse under repeated distillation

*Evidence basis: speculative\.*Two\-timescale systems may risk a degenerative echo: each distillation window compresses skills into the model, and the model then proposes skills resembling what it just absorbed\. No surveyed method \(MetaClaw,SKILL0,K2\-Agent\) runs long enough to test this directly\. The missing evidence is long\-horizon behavior under repeated distillation, especially proposal diversity and dependence between new proposals and recently distilled skills\.

### 12\.5Retrieval beyond moderate library sizes

*Evidence basis: measured at moderate scale, extrapolated to marketplace scale\.*Hierarchy, DAG, and ontology storage push the 64–128\-skill flat\-retrieval drop rightward but do not prove general removal\. The open problem is scaling composition to the 10,000–1,000,000\-skill regime suggested bySkillRouter’s 80K pool,SRA’s 26K\-skill corpus,AgentSkillOS’s 200K\-scale evaluation,SkillNet, andSkilldex\.Graph of Skills\(Liuet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib89)\)improves dependency\-aware retrieval up to 2,000 skills, but marketplace\-scale evidence is missing\. Three directions are ready for comparison: learned routing over skill IDs \(SkillRouter\), skill compilation viaMerge/Distill\(Single\-Agent\-Skills\), and hybrid parametric–retrieval systems that retrieve only when routing confidence is low\.

### 12\.6Cross\-library skill portability and naming

*Evidence basis: measured transfer cases, extrapolated compatibility model\.*Skills increasingly move across agents and users \(SkillClaw,AgentSkillOS,AutoSkill\), but the literature treats portability as deployment rather than learning\. The open problem is*capability compatibility*: when is a skill authored on agentAAsafe and useful on agentBB? Tool API, tokenizer, context window, and harness assumptions can all break transfer\.CoEvoSkillsreports broad cross\-model transfer, whileXSkillreports mixed multimodal transfer; a useful compatibility model should predict such outcomes before evaluation\.

### 12\.7Governance, provenance, and attribution

*Evidence basis: measured attacks and partial defenses, extrapolated integration\.*§[11](https://arxiv.org/html/2607.10113#S11)shows fast growth in attack and measurement work but few integrated defended systems\. The open problem is making ASG\-SI\-style audited graphs, evidence bundles, and verifier\-gated promotion\(Huang and Huang,[2025](https://arxiv.org/html/2607.10113#bib.bib55)\)compatible with10210^\{2\}–10310^\{3\}fast\-loop edits per day and marketplace\-grade triage\.AgentSkills\-Wild,Malicious\-Skills\-Wild,Skill\-Inject,SkillJect,SkillSieve,STARS,Credential Leakage, andMalicious\-or\-Not\(Liuet al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib106);[f](https://arxiv.org/html/2607.10113#bib.bib107); Schmotzet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib108); Jiaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib109); Hou and Yang,[2026](https://arxiv.org/html/2607.10113#bib.bib77); Zhanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib110); Chenet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib92); Holzbaueret al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib95)\)supply pieces of the admission and invocation scanner; missing are operator\-level provenance and costed deployment architectures\. Candidate mechanisms include logical undo over operator logs and cryptographic provenance linking each skill to the trajectory that justified admission\.

### 12\.8Benchmarks that are honest about dynamics

*Evidence basis: measured benchmark gaps\.*SkillFlow\-Benchis the clearest temporal benchmark for discovery, patching, reuse, and repair;Wild\-Skillsis the clearest benchmark for realistic retrieval and curation utility;SWE\-Skills\-Benchis the clearest paired evidence that many public skills add no marginal utility in software engineering;Raw\-Experienceseparates extraction from consumption;SkillsVoteadds skill\-linked credit assignment; andSRAis the clearest decomposition of retrieval, incorporation, and downstream use\. The open problem is to combine temporal patching with retrieval realism while reporting global library trajectory, operator velocity, and usage\-vs\-utility\.

Together, the first four problems improve algorithms inside existing mechanism families, while the last four require community infrastructure for portability, governance, and trajectory\-aware benchmarks\. Method papers and infrastructure papers should therefore be judged together: dynamic skills will not mature if algorithms improve inside isolated libraries while portability, provenance, and benchmark realism lag behind\.

## 13Limitations and Update Policy

This survey draws analytical conclusions about a young area, so its limitations matter\. The corpus is frozen at the May 31, 2026 cutoff in §[2\.3](https://arxiv.org/html/2607.10113#S2.SS3); later work may change the evidence for individual patterns, especially in registry\-scale retrieval and safety\. The taxonomy is therefore an updatable frame, not a final vocabulary\.

The search protocol is broad but not exhaustive\. The corpus was assembled through iterative search and snowballing over a fast\-moving arXiv\-heavy area, so false negatives are likely, especially for unpublished systems, product documentation, non\-English papers, and papers that use tool, memory, workflow, or curriculum terminology without the word “skill”\. The cutoff should therefore be read as an audit boundary over the cited literature, not as a claim that every adjacent artifact\-learning paper has been enumerated\.

The evidence is heterogeneous\. Method papers, benchmarks, infrastructure papers, and safety studies do not support the same kinds of claims: a verifier ablation is stronger evidence for an algorithmic pattern than a single benchmark score, while a registry\-scale measurement is stronger evidence for a deployment surface than for a remedy\. We use the lifecycle framework to align evidence roles rather than rank all papers on one axis\.

The definition of “skill” remains unstable across executable programs, natural\-language lessons, graphs, adapters, memories, and capability labels\. Our six\-sense terminology and skill\-record schema make the boundary explicit, but generic tool\-use papers, memory\-only agents, and fine\-tuning pipelines are excluded unless they produce an externally invocable artifact or evaluate that artifact’s lifecycle\.

The lifecycle framing is also imperfect\. It fits systems that store reusable artifacts outside the model, but purely parametric methods compress proposal, verification, and admission into a training loop; memory\-only methods may retrieve episodes without ever forming an invocable interface; and multimodal or embodied systems may attach verification to a simulator, visual affordance, or physical rollout rather than to a textual artifact\. We include such systems only when they expose enough artifact structure to compare lifecycle choices\.

The operator vocabulary is intentionally coarse\. It does not specify closure, associativity, or edit\-composition laws\. It also does not model every edit dependency, such as aRefinethat triggers downstreamPrune, a graph rewrite that changes both retrieval and composition semantics, or a stochastic proposal distribution that emits many candidates before one is admitted\. The vocabulary is useful for comparing operator repertoires and storage costs, but it is not a full operational semantics for agent development environments\.

Most conclusions are architectural rather than causal\. The evidence supports patterns such as “admission gates matter”, “flat retrieval has a moderate\-size failure mode”, and “benchmarks under\-report library trajectories”, but not a single prescription for the best verifier, storage topology, maintenance schedule, or distillation cadence\. We also do not provide a quantitative cross\-paper leaderboard: backbone, context budget, tool surface, evaluator, and task stream vary too much for pooled effect sizes to be meaningful without a shared harness\.

## 14Broader Impact Statement

Dynamic skill systems can make LLM agents more useful by preserving verified procedures, reducing repeated trial\-and\-error, improving transfer across related tasks, and exposing reusable artifacts for human inspection\. These benefits matter most in operational settings where agents repeatedly use the same tools, interfaces, and domain workflows\.

The same mechanism also creates new risks\. A skill library is a high\-trust execution substrate: admitted artifacts can carry prompt\-injection text, unsafe code, bundled model backdoors, stale procedures, hidden credential leaks, or harmful but well\-formed capabilities\. Registry and marketplace settings add supply\-chain, ownership, attribution, and skill\-stealing concerns\. This survey does not release new agent capabilities, but it organizes design patterns that could be used to build more autonomous skill\-generation pipelines\. The risk is highest if readers adopt growth and sharing mechanisms without the corresponding admission, containment, provenance, and rollback mechanisms analyzed in §[11](https://arxiv.org/html/2607.10113#S11)\.

The mitigation message of the survey is therefore structural\. Dynamic\-skill research should treat safety and governance as lifecycle stages rather than after\-the\-fact filters: candidate skills should pass admission\-time checks, carry manifests for permissions and dependencies, execute inside containment boundaries, retain operator\-level provenance, and support demotion or rollback when downstream evidence changes\. Benchmark papers should report unsafe invocation, skill usage versus skill utility, and maintenance\-off behavior when those quantities are relevant to deployment\.

## 15Conclusion

When agents store reusable procedural knowledge outside the model, the skill library becomes part of the learning system\. It has state, update rules, selection pressure, memory compression, maintenance costs, and safety constraints\. Four survey takeaways are now well supported; one scaling boundary remains open\.

The first takeaway is*the lifecycle view*\. Dynamic skill systems are lifecycle\-managed, verified, evolving artifact stores\. The decomposition in §[5](https://arxiv.org/html/2607.10113#S5)makesSkillWeaver,PSN,SkillRouter,MetaClaw,SkillFoundry,Skilldex,ClawSafety,SkillFlow\-Bench,SkillOpt, andSkillsVotecomparable without pretending that they solve the same problem\.

The second takeaway is*the time index*\. Viewing libraries as dynamic objectsℒt\\mathcal\{L\}\_\{t\}makes maintenance\-off ablations, library\-size drops, verifier\-quality sensitivity, usage\-vs\-utility gaps, and two\-timescale decoupling expressible\. The skill\-record schema⟨C,π,T,R,φ,ν,≺⟩\\langle C,\\pi,T,R,\\varphi,\\nu,\\prec\\rangleprovides the structural language for those analyses\.

The third takeaway is*write\-time discipline*\. Admission, verifier quality under RL, maintenance, and write\-time abstraction all concern what enters the library, under what gate, and with what abstraction\. The newest skill\-optimization papers make the point explicit: the editable skill artifact is the trainable external state, and validation decides which textual updates survive\. The evidence aligns with a broader lifelong\-learning principle: what a system keeps can matter more than what it sees\.

The fourth takeaway is*mechanism pluralism*\. The operator vocabulary describes multiple coherent specializations: retrospective induction, execution\-verified skill writing, skill\-aware RL, cross\-user aggregation, graph/hierarchy management, and two\-timescale internalization\. These are not separate taxonomies; they are different ways of coupling edit repertoire, admission, storage, and update clocks\.

The open boundary is scaling\. Moderate\-library\-size retrieval degradation is recurrent; hierarchy, DAG, ontology, routing, and orchestration push it rightward but have not shown general removal\.SkillFlow\-Benchimproves evaluation by tracking discovery, patching, repair, skill count, and usage, but its family\-local protocol does not answer global heterogeneous\-library scaling\. Retrieval scaling, cross\-library portability, provenance, and benchmark honesty remain open\.

#### Reporting checklist\.

For future work to become comparable, papers should report five quantities:

1. 1\.which lifecycle stages are implemented and which are only assumed;
2. 2\.operator velocities and library trajectories over time, not only final task success;
3. 3\.repair\-, maintenance\-, or admission\-off ablations when those stages are part of the method;
4. 4\.skill usage separately from skill utility;
5. 5\.the operators that an infrastructure or safety substrate makes cheap, expensive, blocked, or auditable\.

Safety papers should additionally evaluate admission, composition, and provenance surfaces rather than treating skills as ordinary prompt files\. If these quantities become standard, the next survey can make quantitative cross\-method comparisons that this one deliberately avoids\.

## Acknowledgments

## References

- E\. C\. Acikgoz, C\. Qian, J\. Hübotter, H\. Ji, D\. Hakkani\-Tür, and G\. Tur \(2026\)Tool\-R0: Self\-Evolving LLM Agents for Tool\-Learning from Zero Data\.arXiv:2602\.21320\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.72.66.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.5.4.1)\.
- M\. Allard, A\. Teinturier, V\. Xing, and G\. Viaud \(2026\)Experiential Reflective Learning for Self\-Improving LLM Agents\.arXiv:2603\.24639\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.19.13.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.32.2.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1)\.
- S\. Alzubi, N\. Provenzano, J\. Bingham, W\. Chen, and T\. Vu \(2026\)EvoSkill: Automated Skill Discovery for Multi\-Agent Systems\.arXiv:2603\.02766\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.35.29.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- Anthropic \(2025\)Agent Skills: Specification and Documentation\.Note:GitHub repositoryOpen format/specification repository; accessed May 17, 2026External Links:[Link](https://github.com/agentskills/agentskills)Cited by:[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- P\. Bacon, J\. Harb, and D\. Precup \(2017\)The Option\-Critic Architecture\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.31,pp\. 1726–1734\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v31i1.10916)Cited by:[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p2.1)\.
- S\. Bi, M\. Wu, H\. Hao, K\. Li, W\. Liu, S\. Song, H\. Zhao, and A\. Zhou \(2026\)Automating Skill Acquisition through Large\-Scale Mining of Open\-Source Agentic Repositories: A Framework for Multi\-Agent Procedural Knowledge Extraction\.arXiv:2603\.11808\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.94.88.1)\.
- R\. Bornemann, P\. V\. Amadori, and A\. Cully \(2026\)CODE\-SHARP: Continuous Open\-ended Discovery and Evolution of Skills as Hierarchical Reward Programs\.arXiv:2602\.10085\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.13.7.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1)\.
- T\. Cai, X\. Wang, T\. Ma, X\. Chen, and D\. Zhou \(2023\)Large Language Models as Tool Makers\.arXiv:2305\.17126\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.8.2.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.2.2.2.2),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1)\.
- Y\. Cai, Y\. Hao, J\. Zhou, H\. Yan, Z\. Lei, R\. Zhen, Z\. Han, Y\. Yang, J\. Li, Q\. Pan, T\. Huai, Q\. Chen, X\. Li, K\. Chen, B\. Zhang, X\. Qiu, and L\. He \(2025\)Building Self\-Evolving Agents via Experience\-Driven Lifelong Learning: A Framework and Benchmark\.arXiv:2508\.19005\.Cited by:[§2\.1](https://arxiv.org/html/2607.10113#S2.SS1.p1.1),[§8\.3](https://arxiv.org/html/2607.10113#S8.SS3.p1.1)\.
- L\. Chen, E\. Feng, Y\. Xia, and H\. Chen \(2026a\)SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses\.arXiv:2604\.03088\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.79.73.1)\.
- Q\. Chen, A\. Bellucci, Z\. Sun, and G\. Jacucci \(2026b\)SkillDroid: Compile Once, Reuse Forever\.arXiv:2604\.14872\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.33.27.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- S\. Chen, J\. Gai, R\. Zhou, J\. Zhang, T\. Zhu, J\. Li, K\. Wang, Z\. Wang, Z\. Chen, K\. Kaleb, N\. Miao, S\. Gao, C\. Lu, M\. Li, J\. He, and Y\. W\. Teh \(2026c\)SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?\.arXiv:2603\.00718\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.38.32.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2)\.
- T\. Chen, Y\. Li, M\. Solodko, S\. Wang, N\. Jiang, T\. Cui, J\. Hao, J\. Ko, S\. Abdali, L\. Xu, S\. Zheng, H\. Fan, P\. Cameron, J\. Wagle, and K\. Koishida \(2026d\)CUA\-Skill: Develop Skills for Computer Using Agent\.arXiv:2601\.21123\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.48.42.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1)\.
- X\. Chen, C\. Xu, Y\. Wang, B\. Liu, Z\. Yao, and Y\. He \(2026e\)Learning to Self\-Evolve\.arXiv:2603\.18620\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.66.60.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.21.21.21.4),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.12.11.1)\.
- Z\. Chen, Y\. Zhang, Y\. Liu, G\. Deng, Y\. Li, Y\. Zhang, J\. Ning, L\. Zhang, L\. Ma, and Z\. Li \(2026f\)Credential Leakage in LLM Agent Skills: A Large\-Scale Empirical Study\.arXiv:2604\.03070\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.115.109.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px4.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- Z\. Duan, Y\. Tian, Z\. Yin, L\. Pang, J\. Deng, Z\. Wei, S\. Xu, Y\. Ge, and X\. Cheng \(2026\)SkillAttack: Automated Red Teaming of Agent Skills through Attack Path Refinement\.arXiv:2604\.04989\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.113.107.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px1.p1.1)\.
- B\. Eysenbach, A\. Gupta, J\. Ibarz, and S\. Levine \(2019\)Diversity is All You Need: Learning Skills without a Reward Function\.InInternational Conference on Learning Representations,Cited by:[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p2.1)\.
- L\. Fan, P\. Dai, Z\. Deng, H\. Wang, X\. Gong, Y\. Zheng, and Y\. Ou \(2026\)Evolving Medical Imaging Agents via Experience\-driven Self\-skill Discovery\.arXiv:2603\.05860\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.56.50.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1)\.
- J\. Fang, Y\. Peng, X\. Zhang, Y\. Wang, X\. Yi, G\. Zhang, Y\. Xu, B\. Wu, S\. Liu, Z\. Li, Z\. Ren, N\. Aletras, X\. Wang, H\. Zhou, and Z\. Meng \(2025\)A Comprehensive Survey of Self\-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems\.arXiv:2508\.07407\.Cited by:[§1\.2](https://arxiv.org/html/2607.10113#S1.SS2.p3.1),[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p1.1)\.
- L\. Gaven, T\. Carta, C\. Romac, C\. Colas, S\. Lamprier, O\. Sigaud, and P\. Oudeyer \(2025\)MAGELLAN: Metacognitive Predictions of Learning Progress Guide Autotelic LLM Agents in Large Goal Spaces\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 18920–18953\.Cited by:[§2\.1](https://arxiv.org/html/2607.10113#S2.SS1.p1.1)\.
- J\. Gong, R\. Gu, Z\. Fei, Y\. Cao, L\. Twist, A\. Geiger, S\. Han, D\. Sobania, F\. Sarro, and J\. M\. Zhang \(2026\)SkillMOO: Multi\-Objective Optimization of Agent Skills for Software Engineering\.arXiv:2604\.09297\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.52.46.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- T\. Han, Y\. Zhang, W\. Song, C\. Fang, Z\. Chen, Y\. Sun, and L\. Hu \(2026\)SWE\-Skills\-Bench: Do Agent Skills Actually Help in Real\-World Software Engineering?\.arXiv:2603\.15401\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.84.78.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- F\. Holzbauer, D\. Schmidt, G\. Gegenhuber, S\. Schrittwieser, and J\. Ullrich \(2026\)Malicious Or Not: Adding Repository Context to Agent Skill Classification\.arXiv:2603\.16572\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.118.112.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px5.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- Y\. Hou and Z\. Yang \(2026\)SkillSieve: A Hierarchical Triage Framework for Detecting Malicious AI Agent Skills\.arXiv:2604\.06550\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.112.106.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px5.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- Y\. Hou, X\. Lao, H\. Wang, Q\. Yao, W\. Chen, B\. Huang, F\. Sun, Y\. Lv, W\. Lei, X\. Wen, S\. Xie, P\. Xia, and Z\. Tan \(2026\)MedSkillAudit: A Domain\-Specific Audit Framework for Medical Research Agent Skills\.arXiv:2604\.20441\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.120.114.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px7.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p3.1)\.
- C\. Huang, H\. Zhang, J\. Xu, Z\. Zheng, and Y\. Lin \(2026a\)Bilevel Optimization of Agent Skills via Monte Carlo Tree Search\.arXiv:2604\.15709\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.53.47.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1)\.
- K\. Huang and J\. Huang \(2025\)Audited Skill\-Graph Self\-Improvement for Agentic LLMs via Verifiable Rewards, Experience Synthesis, and Continual Memory\.arXiv:2512\.23760\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.103.97.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px8.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.10.9.1),[§9\.2](https://arxiv.org/html/2607.10113#S9.SS2.p1.5)\.
- X\. Huang, J\. Chen, Y\. Fei, Z\. Li, P\. Schwaller, and G\. Ceder \(2025\)CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution\.arXiv:2512\.23880\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.40.34.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.24.24.24.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- Y\. Huang, B\. Zheng, Y\. Dong, H\. Tang, H\. Zhao, S\. M\. R\. H\. Shawon, and H\. Zhang \(2026b\)A Self\-Evolving Agentic Framework for Metasurface Inverse Design\.arXiv:2604\.01480\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.54.48.1)\.
- Z\. Huang, J\. Xu, Y\. Yang, Z\. Gong, Q\. Yang, M\. Tian, X\. Wang, C\. Lv, X\. Gao, Q\. Dai, B\. Liu, K\. Qiu, X\. Yang, D\. Chen, X\. Zheng, and C\. Luo \(2026c\)From Raw Experience to Skill Consumption: A Systematic Study of Model\-Generated Agent Skills\.arXiv:2605\.23899\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.97.91.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p3.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- X\. Jia, J\. Liao, S\. Qin, J\. Gu, W\. Ren, X\. Cao, Y\. Liu, and P\. Torr \(2026\)SkillJect: Automating Stealthy Skill\-Based Prompt Injection for Coding Agents with Trace\-Driven Closed\-Loop Refinement\.arXiv:2602\.14211\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.111.105.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px1.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- G\. Jiang, Z\. Su, X\. Qu, and Y\. R\. Fung \(2026a\)XSkill: Continual Learning from Experience and Skills in Multimodal Agents\.arXiv:2603\.12056\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.24.18.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.25.25.25.2),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§8\.2](https://arxiv.org/html/2607.10113#S8.SS2.p2.4),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- Y\. Jiang, D\. Li, H\. Deng, B\. Ma, X\. Wang, Q\. Wang, and G\. Yu \(2026b\)SoK: Agentic Skills – Beyond Tool Use in LLM Agents\.arXiv:2602\.20867\.Cited by:[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.10113#S2.SS1.p1.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px1.p1.4)\.
- Y\. Jiang, Y\. Zhang, M\. Backes, X\. Shen, and Y\. Zhang \(2026c\)HarmfulSkillBench: How Do Harmful Skills Weaponize Your Agents?\.arXiv:2604\.15415\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.116.110.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px3.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p3.1)\.
- Z\. Jiao, S\. Wang, Z\. Zhang, X\. Ren, W\. Wang, B\. Zhao, H\. Wei, and L\. Zhang \(2026\)Agentic Proposing: Enhancing Large Language Model Reasoning via Compositional Skill Synthesis\.arXiv:2602\.03279\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.14.8.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.14.13.1)\.
- R\. Ju, X\. Wang, X\. Ding, Y\. Yang, H\. Wu, S\. Jiang, Q\. Zhang, H\. Wen, X\. Li, W\. Wang, K\. Li, Y\. Liu, H\. Dai, W\. Wang, and T\. Cao \(2026\)EmbodiSkill: Skill\-Aware Reflection for Self\-Evolving Embodied Agents\.arXiv:2605\.10332\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.26.20.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- Y\. Jung, T\. Padhi, S\. Shaham, D\. Khullar, J\. Jeong, N\. Mehrabi, and E\. Yang \(2025\)Co\-Evolving Agents: Learning from Failures as Hard Negatives\.arXiv:2511\.22254\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.70.64.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.29.29.29.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.13.12.1),[§9\.2](https://arxiv.org/html/2607.10113#S9.SS2.p1.5)\.
- G\. Konidaris and A\. G\. Barto \(2009\)Skill Discovery in Continuous Reinforcement Learning Domains Using Skill Chaining\.InAdvances in Neural Information Processing Systems,Vol\.22,pp\. 1015–1023\.Cited by:[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p2.1)\.
- F\. Li, P\. Tagkopoulos, and I\. Tagkopoulos \(2025\)SkillFlow: Scalable and Efficient Agent Skill Retrieval System\.arXiv:2504\.06188\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.82.76.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.26.26.26.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1)\.
- G\. Li, Y\. Xie, Y\. Liu, Z\. Dong, X\. Pan, T\. Zheng, J\. Choi, M\. Morais, B\. Jha, S\. Mishra, B\. Zhou, C\. Luo, M\. Cheng, and D\. Song \(2026a\)The World Won’t Stay Still: Programmable Evolution for Agent Benchmarks\.arXiv:2603\.05910\.Cited by:[§8\.3](https://arxiv.org/html/2607.10113#S8.SS3.p1.1)\.
- H\. Li, C\. Mu, J\. Chen, S\. Ren, Z\. Cui, Y\. Zhang, L\. Bai, and S\. Hu \(2026b\)Organizing, Orchestrating, and Benchmarking Agent Skills at Ecosystem Scale\.arXiv:2603\.02176\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.77.71.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§4](https://arxiv.org/html/2607.10113#S4.p4.3),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- X\. Li, Y\. Liu, W\. Chen, S\. Zheng,et al\.\(2026c\)SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks\.arXiv:2602\.12670\.Cited by:[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5)\.
- X\. Li \(2026\)When Single\-Agent with Skills Replace Multi\-Agent Systems and When They Fail\.arXiv:2601\.04748\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.3.3.2),[§4](https://arxiv.org/html/2607.10113#S4.p4.3),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- Y\. Li, W\. Liu, B\. Gao, J\. Xie, Z\. Ai, N\. Zou, Y\. Li, and T\. Fu \(2026d\)SkillsInjector: Dynamic Skill Context Construction for LLM Agents\.arXiv:2605\.29794\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.101.95.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p1.1),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- Y\. Li, R\. Miao, Z\. Qi, and T\. Lan \(2026e\)ARISE: Agent Reasoning with Intrinsic Skill Evolution in Hierarchical Reinforcement Learning\.arXiv:2603\.16060\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.74.68.1)\.
- Z\. Li, J\. Wu, X\. Ling, X\. Cui, and T\. Luo \(2026f\)Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis\.arXiv:2604\.02837\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.105.99.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.p1.1)\.
- Q\. Liang, H\. Wang, Z\. Liang, and Y\. Liu \(2026a\)From Skill Text to Skill Structure: The Scheduling\-Structural\-Logical Representation for Agent Skills\.arXiv:2604\.24026\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.5.3),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.38.8.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p2.1)\.
- Y\. Liang, R\. Zhong, H\. Xu, C\. Jiang, Y\. Zhong, R\. Fang, J\. Gu, S\. Deng, Y\. Yao, M\. Wang, S\. Qiao, X\. Xu, T\. Wu, K\. Wang, Y\. Liu, Z\. Bi, J\. Lou, Y\. E\. Jiang, H\. Zhu, G\. Yu, H\. Hong, L\. Huang, H\. Xue, C\. Wang, Y\. Wang, Z\. Shan, X\. Chen, Z\. Tu, F\. Xiong, X\. Xie, P\. Zhang, Z\. Gui, L\. Liang, J\. Zhou, C\. Wu, J\. Shang, Y\. Gong, J\. Lin, C\. Xu, H\. Deng, W\. Zhang, K\. Ding, Q\. Zhang, F\. Huang, N\. Zhang, J\. Z\. Pan, G\. Qi, H\. Wang, and H\. Chen \(2026b\)SkillNet: Create, Evaluate, and Connect AI Skills\.arXiv:2603\.04448\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.80.74.1)\.
- H\. Lin, P\. Li, J\. Song, F\. Jiang, and T\. Zhang \(2026\)MUSE\-Autoskill: Self\-Evolving Agents via Skill Creation, Memory, Management, and Evaluation\.arXiv:2605\.27366\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.61.55.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.15.15.15.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- G\. Ling, S\. Zhong, and R\. Huang \(2026\)Agent Skills: A Data\-Driven Analysis of Claude Skills for Extending Large Language Model Functionality\.arXiv:2602\.08004\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.95.89.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1)\.
- D\. Liu, Z\. Li, H\. Du, X\. Wu, S\. Gui, Y\. Kuang, and L\. Sun \(2026a\)Graph of Skills: Dependency\-Aware Structural Retrieval for Massive Agent Skills\.arXiv:2604\.05333\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.90.84.1),[§12\.5](https://arxiv.org/html/2607.10113#S12.SS5.p1.2),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- H\. Liu, Y\. Ming, S\. Joty, and C\. Zhao \(2026b\)Harnessing LLM Agents with Skill Programs\.arXiv:2605\.17734\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.59.53.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.7.7.7.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- H\. Liu, H\. Yang, T\. Jiang, B\. Tang, F\. Xiong, and Z\. Li \(2026c\)SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution\.arXiv:2605\.18401\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.98.92.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.28.28.28.2),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p2.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- J\. Liu, Y\. Su, P\. Xia, S\. Han, Z\. Zheng, C\. Xie, M\. Ding, and H\. Yao \(2026d\)SimpleMem: Efficient Lifelong Memory for LLM Agents\.arXiv:2601\.02553\.Cited by:[Table 1](https://arxiv.org/html/2607.10113#S1.T1.22.22.22.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- X\. Liu, X\. Luo, L\. Li, G\. Huang, J\. Liu, and H\. Qiao \(2026e\)SkillForge: Forging Domain\-Specific, Self\-Evolving Agent Skills in Cloud Technical Support\.arXiv:2604\.08618\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.51.45.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- Y\. Liu, Z\. Chen, Y\. Zhang, G\. Deng, Y\. Li, J\. Ning, Y\. Zhang, and L\. Y\. Zhang \(2026f\)Malicious Agent Skills in the Wild: A Large\-Scale Security Empirical Study\.arXiv:2602\.06547\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.109.103.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px2.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- Y\. Liu, W\. Wang, R\. Feng, Y\. Zhang, G\. Xu, G\. Deng, Y\. Li, and L\. Zhang \(2026g\)Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale\.arXiv:2601\.10338\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.108.102.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px2.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- Y\. Liu, J\. Ji, L\. An, T\. Jaakkola, Y\. Zhang, and S\. Chang \(2026h\)How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings\.arXiv:2604\.04323\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.47.41.1),[§4](https://arxiv.org/html/2607.10113#S4.p4.3),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- Z\. Lu, Z\. Yao, J\. Wu, C\. Han, Q\. Gu, X\. Cai, W\. Lu, J\. Xiao, Y\. Zhuang, and Y\. Shen \(2026a\)SKILL0: In\-Context Agentic Reinforcement Learning for Skill Internalization\.arXiv:2604\.02268\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.65.59.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.17.17.17.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p5.4),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- Z\. Lu, Y\. Zuo, Y\. Nie, X\. He, W\. Fan, L\. Qi, and S\. Jin \(2026b\)ContractSkill: Repairable Contract\-Based Skills for Multimodal Web Agents\.arXiv:2603\.20340\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.32.26.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.36.6.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- Y\. Ma, Y\. Huang, H\. Bao, H\. Zhuang, S\. Shukla, M\. Galley, X\. Zhang, and S\. Feuerriegel \(2026a\)SkillGen: Verified Inference\-Time Agent Skill Synthesis\.arXiv:2605\.10999\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.58.52.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p3.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p3.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- Z\. Ma, S\. Yang, Y\. Ji, X\. Wang, Y\. Wang, Y\. Hu, T\. Huang, and X\. Chu \(2026b\)SkillClaw: Let Skills Evolve Collectively with Agentic Evolver\.arXiv:2604\.08377\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.43.37.1),[§1](https://arxiv.org/html/2607.10113#S1.p4.1)\.
- O\. Nachum, S\. Gu, H\. Lee, and S\. Levine \(2018\)Data\-Efficient Hierarchical Reinforcement Learning\.InAdvances in Neural Information Processing Systems,Vol\.31,pp\. 3307–3317\.Cited by:[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p2.1)\.
- J\. Ni, Y\. Liu, X\. Liu, Y\. Sun, M\. Zhou, P\. Cheng, D\. Wang, E\. Zhao, X\. Jiang, and G\. Jiang \(2026\)Trace2Skill: Distill Trajectory\-Local Lessons into Transferable Agent Skills\.arXiv:2603\.25158\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.42.36.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.11.11.11.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- Z\. Nie, R\. Shen, X\. Yu, B\. Yin, J\. Zhang, and X\. Hu \(2026\)SkillGraph: Self\-Evolving Multi\-Agent Collaboration with Multimodal Graph Topology\.arXiv:2604\.17503\.Cited by:[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p2.1)\.
- L\. Qiu, Z\. Gao, J\. Chen, Y\. Ye, W\. Huang, X\. Xue, W\. Qiu, and S\. Tang \(2026\)AutoRefine: From Trajectories to Reusable Expertise for Continual LLM Agent Refinement\.arXiv:2601\.22758\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.45.39.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.12.12.12.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§3\.4](https://arxiv.org/html/2607.10113#S3.SS4.p1.17),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- Y\. Qu, Y\. Liu, T\. Geng, G\. Deng, Y\. Li, L\. Zhang, Y\. Zhang, and L\. Ma \(2026\)Supply\-Chain Poisoning Attacks Against LLM Coding Agent Skill Ecosystems\.arXiv:2604\.03081\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.107.101.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px2.p1.1)\.
- S\. Saha and P\. Hemanth \(2026\)Skilldex: A Package Manager and Registry for Agent Skill Packages with Hierarchical Scope\-Based Distribution\.arXiv:2604\.16911\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.92.86.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1)\.
- S\. Saha, K\. Faghih, and S\. Feizi \(2026\)Under the Hood of SKILL\.md: Semantic Supply\-chain Attacks on AI Agent Skill Registry\.arXiv:2605\.11418\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.121.115.1),[§10\.3](https://arxiv.org/html/2607.10113#S10.SS3.p1.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px2.p1.1)\.
- D\. Schmotz, S\. Abdelnabi, and M\. Andriushchenko \(2025\)Agent Skills Enable a New Class of Realistic and Trivially Simple Prompt Injections\.arXiv:2510\.26328\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.106.100.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px1.p1.1)\.
- D\. Schmotz, L\. Beurer\-Kellner, S\. Abdelnabi, and M\. Andriushchenko \(2026\)Skill\-Inject: Measuring Agent Vulnerability to Skill File Attacks\.arXiv:2602\.20156\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.110.104.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px1.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- S\. Shen, W\. Cheng, M\. Ma, A\. Turcan, M\. J\. Zhang, and J\. Ma \(2026\)SkillFoundry: Building Self\-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources\.arXiv:2604\.03964\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.50.44.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- H\. Shi, X\. Yuan, and B\. Liu \(2026\)Evolving Programmatic Skill Networks\.arXiv:2601\.03509\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.37.31.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1)\.
- W\. Song, J\. Yue, and Z\. Pang \(2026a\)ABSTRAL: Automatic Design of Multi\-Agent Systems Through Iterative Refinement and Topology Optimization\.arXiv:2603\.22791\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.49.43.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.37.7.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- X\. Song, H\. Pu, and L\. Zhao \(2026b\)SkillOps: Managing LLM Agent Skill Libraries as Self\-Maintaining Software Ecosystems\.arXiv:2605\.13716\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.100.94.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.8.8.8.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p2.3),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- W\. Su, J\. Long, Q\. Ai, Y\. Tang, C\. Wang, Y\. Tu, and Y\. Liu \(2026\)Skill Retrieval Augmentation for Agentic AI\.arXiv:2604\.24594\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.85.79.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.27.27.27.2),[§10\.2](https://arxiv.org/html/2607.10113#S10.SS2.p1.1),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- Y\. Sun, P\. Wei, and L\. B\. Hsieh \(2026\)Don’t Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG\.arXiv:2604\.14572\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.93.87.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1)\.
- R\. S\. Sutton, D\. Precup, and S\. Singh \(1999\)Between MDPs and Semi\-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning\.Artificial Intelligence112\(1–2\),pp\. 181–211\.External Links:[Document](https://dx.doi.org/10.1016/S0004-3702%2899%2900052-1)Cited by:[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p2.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px1.p1.4)\.
- Z\. Tao, R\. Lai, C\. Yu, W\. Chen, L\. Chen, B\. Kong, L\. Cheng, C\. Zhuo, Z\. Li, and Q\. Sun \(2026\)SAGER: Self\-Evolving User Policy Skills for Recommendation Agent\.arXiv:2604\.14972\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.25.19.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1)\.
- G\. Tie, J\. Shi, P\. Zhou, and L\. Sun \(2026\)BadSkill: Backdoor Attacks on Agent Skills via Model\-in\-Skill Poisoning\.arXiv:2604\.09378\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.114.108.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px2.p1.1)\.
- C\. Wang, Z\. Yu, X\. Xie, W\. Yao, R\. Fang, S\. Qiao, K\. Cao, G\. Zheng, X\. Qi, P\. Zhang, and S\. Deng \(2026a\)SkillX: Automatically Constructing Skill Knowledge Bases for Agents\.arXiv:2604\.04804\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.41.35.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar \(2023\)Voyager: An Open\-Ended Embodied Agent with Large Language Models\.arXiv:2305\.16291\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.2.2.3),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.1.1.1.2),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1)\.
- H\. Wang, Y\. Lan, B\. Cao, L\. Lin, and J\. Chen \(2026b\)SkillGrad: Optimizing Agent Skills Like Gradient Descent\.arXiv:2605\.27760\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.62.56.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.14.14.14.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p3.1),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- J\. Wang, Y\. Ming, Z\. Ke, S\. Joty, A\. Albarghouthi, and F\. Sala \(2026c\)SkillOrchestra: Learning to Route Agents via Skill Transfer\.arXiv:2602\.19672\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.81.75.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px8.p1.1),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.SSS0.Px1.p1.1)\.
- J\. Wang, C\. Zhou, Z\. Fu, J\. Wang, W\. Liu, W\. Zhang, and J\. Lin \(2026d\)Skills on the Fly: Test\-Time Adaptive Skill Synthesis for LLM Agents\.arXiv:2605\.16986\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.27.21.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.SSS0.Px1.p1.1)\.
- J\. Wang, Q\. Yan, Y\. Wang, Y\. Tian, S\. S\. Mishra, Z\. Xu, M\. Gandhi, P\. Xu, and L\. L\. Cheong \(2025a\)Reinforcement Learning for Self\-Improving Agent with Skill Library\.arXiv:2512\.17102\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.10.4.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.5.5.5.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.6.5.1)\.
- X\. Wang, N\. Liao, S\. Wei, C\. Tang, and F\. Xiong \(2026e\)AutoAgent: Evolving Cognition and Elastic Memory Orchestration for Adaptive Agents\.arXiv:2603\.09716\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.86.80.1)\.
- X\. Wang, Z\. Han, Z\. Liu, G\. Li, J\. Dong, B\. Liu, L\. Liu, and Z\. Han \(2026f\)Lifelong Language\-Conditioned Robotic Manipulation Learning\.arXiv:2603\.05160\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.67.61.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.16.16.16.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p5.4)\.
- Z\. Wang, Q\. Wu, X\. Zhang, C\. Zhang, W\. Yao, F\. E\. Faisal, B\. Peng, S\. Qin, S\. Nath, Q\. Lin, C\. Bansal, D\. Zhang, S\. Rajmohan, J\. Gao, and H\. Yao \(2026g\)WebXSkill: Skill Learning for Autonomous Web Agents\.arXiv:2604\.13318\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.31.25.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- Z\. Wang, R\. Zhang, Y\. Liu, C\. Liu, Q\. Zhao, H\. Li, and G\. Xu \(2026h\)Black\-Box Skill Stealing Attack from Proprietary LLM Agents: An Empirical Study\.arXiv:2604\.21829\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.117.111.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px6.p1.1)\.
- Z\. Wang, Y\. Shi, M\. Li, Z\. Liu, J\. M\. Zhang, C\. Wan, and X\. Gu \(2026i\)EffiSkill: Agent Skill Based Automated Code Efficiency Optimization\.arXiv:2603\.27850\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.89.83.1),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12)\.
- Z\. Z\. Wang, A\. Gandhi, G\. Neubig, and D\. Fried \(2025b\)Inducing Programmatic Skills for Agentic Tasks\.arXiv:2504\.06821\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.30.24.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.6.6.6.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- B\. Wei, Y\. Zhang, J\. Pan, K\. Mei, X\. Wang, J\. Hamm, Z\. Zhu, and Y\. Ge \(2026\)ClawSafety: "Safe" LLMs, Unsafe Agents\.arXiv:2604\.01438\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.104.98.1),[§1](https://arxiv.org/html/2607.10113#S1.p4.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px1.p1.1),[§3\.3](https://arxiv.org/html/2607.10113#S3.SS3.p1.24)\.
- Z\. A\. Wen \(2026\)Toward User Comprehension Supports for LLM Agent Skill Specifications\.arXiv:2605\.19362\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.122.116.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px5.p1.1)\.
- Z\. Weng, A\. Antoniades, D\. Nathani, Z\. Zhang, X\. Pu, and X\. E\. Wang \(2026\)Group\-Evolving Agents: Open\-Ended Self\-Improvement via Experience Sharing\.arXiv:2602\.04837\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.88.82.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.30.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1)\.
- R\. Wu, X\. Wang, J\. Mei, P\. Cai, D\. Fu, C\. Yang, L\. Wen, X\. Yang, Y\. Shen, Y\. Wang, and B\. Shi \(2025\)EvolveR: Self\-Evolving LLM Agents through an Experience\-Driven Lifecycle\.arXiv:2510\.16079\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.22.16.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.34.4.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- X\. Wu, Z\. Li, G\. Shi, A\. Duffy, T\. Marques, M\. L\. Olson, T\. Zhou, and D\. Manocha \(2026a\)Co\-Evolving LLM Decision and Skill Bank Agents for Long\-Horizon Tasks\.arXiv:2604\.20987\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.15.9.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.7.6.1)\.
- Z\. Wu, D\. Mo, H\. Lu, J\. Xing, J\. Liu, Y\. Jing, K\. Li, K\. Shao, J\. Hao, and Y\. Shi \(2026b\)K2\-Agent: Co\-Evolving Know\-What and Know\-How for Hierarchical Mobile Device Control\.arXiv:2603\.00676\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.68.62.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p5.4),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.1.2),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- C\. S\. Xia, Z\. Wang, Y\. Yang, Y\. Wei, and L\. Zhang \(2025a\)Live\-SWE\-agent: Can Software Engineering Agents Self\-Evolve on the Fly?\.arXiv:2511\.13646\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.39.33.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.4.4.4.2),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- P\. Xia, J\. Chen, H\. Wang, J\. Liu, K\. Zeng, Y\. Wang, S\. Han, Y\. Zhou, X\. Zhao, H\. Chen, Z\. Zheng, C\. Xie, and H\. Yao \(2026a\)SkillRL: Evolving Agents via Recursive Skill\-Augmented Reinforcement Learning\.arXiv:2602\.08234\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.11.5.1)\.
- P\. Xia, J\. Chen, X\. Yang, H\. Tu, J\. Liu, K\. Xiong, S\. Han, S\. Qiu, H\. Ji, Y\. Zhou, Z\. Zheng, C\. Xie, and H\. Yao \(2026b\)MetaClaw: Just Talk – An Agent That Meta\-Learns and Evolves in the Wild\.arXiv:2603\.17187\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.69.63.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.9.9.9.2),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p5.4),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.3.2.1),[§9\.4](https://arxiv.org/html/2607.10113#S9.SS4.p1.1)\.
- P\. Xia, K\. Zeng, J\. Liu, C\. Qin, F\. Wu, Y\. Zhou, C\. Xiong, and H\. Yao \(2025b\)Agent0: Unleashing Self\-Evolving Agents from Zero Data via Tool\-Integrated Reasoning\.arXiv:2511\.16043\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.71.65.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.8.7.1)\.
- T\. Xia, L\. Hu, Y\. Sun, M\. Xu, L\. Xu, S\. Wang, W\. Xu, and J\. Jiang \(2026c\)GraSP: Graph\-Structured Skill Compositions for LLM Agents\.arXiv:2604\.17870\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.91.85.1)\.
- S\. Xie, Y\. Zhang, R\. Wang, and X\. Chen \(2026\)Uni\-Skill: Building Self\-Evolving Skill Repository for Generalizable Robotic Manipulation\.arXiv:2603\.02623\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.57.51.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1)\.
- D\. Xu, Z\. Chen, Z\. Pan, J\. Guan, D\. Dong, J\. Li, and B\. Pu \(2026\)SkillSmith: Compiling Agent Skills into Boundary\-Guided Runtime Interfaces\.arXiv:2605\.15215\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.99.93.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p2.1)\.
- R\. Xu and Y\. Yan \(2026\)Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward\.arXiv:2602\.12430\.Cited by:[§1\.2](https://arxiv.org/html/2607.10113#S1.SS2.p3.1),[§2\.1](https://arxiv.org/html/2607.10113#S2.SS1.p1.1),[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p1.1)\.
- C\. Yang, X\. Yang, L\. Wen, D\. Fu, J\. Mei, R\. Wu, P\. Cai, Y\. Shen, N\. Deng, B\. Shi, Y\. Qiao, and H\. Li \(2025a\)Learning on the Job: An Experience\-Driven Self\-Evolving Agent for Long\-Horizon Tasks\.arXiv:2510\.08002\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.21.15.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.23.23.23.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.SSS0.Px1.p1.1)\.
- M\. Yang, J\. Piao, X\. Xia, X\. Lan, J\. Chen, Y\. Gong, and Y\. Li \(2026a\)SkillMaster: Toward Autonomous Skill Mastery in LLM Agents\.arXiv:2605\.08693\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.16.10.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.2](https://arxiv.org/html/2607.10113#S9.SS2.p1.5)\.
- Y\. Yang, Z\. Gong, W\. Huang, Q\. Yang, Z\. Zhou, Z\. Huang, Y\. Li, X\. Gao, Q\. Dai, B\. Liu, K\. Qiu, Y\. Yang, D\. Chen, X\. Yang, and C\. Luo \(2026b\)SkillOpt: Executive Strategy for Self\-Evolving Agent Skills\.arXiv:2605\.23904\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.60.54.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.35.5.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p3.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.6](https://arxiv.org/html/2607.10113#S9.SS6.p1.12),[§9\.7](https://arxiv.org/html/2607.10113#S9.SS7.p1.7)\.
- Y\. Yang, S\. Kang, J\. Lee, D\. Lee, S\. Yun, and K\. Lee \(2025b\)Automated Skill Discovery for Language Agents through Exploration and Iterative Feedback\.arXiv:2506\.04287\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.75.69.1)\.
- Y\. Yang, J\. Li, Q\. Pan, B\. Zhan, Y\. Cai, L\. Du, J\. Zhou, K\. Chen, Q\. Chen, X\. Li, B\. Zhang, and L\. He \(2026c\)AutoSkill: Experience\-Driven Lifelong Learning via Skill Self\-Evolution\.arXiv:2603\.01145\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.44.38.1)\.
- R\. Zabounidis, Y\. Wu, S\. Stepputtis, W\. Kim, Y\. Li, T\. Mitchell, and K\. Sycara \(2026\)SCALAR: Learning and Composing Skills through LLM Guided Symbolic Planning and Deep RL Grounding\.arXiv:2603\.09036\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.73.67.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.9.8.1)\.
- Y\. Zhai, S\. Tao, C\. Chen, A\. Zou, Z\. Chen, Q\. Fu, S\. Mai, L\. Yu, J\. Deng, Z\. Cao, Z\. Liu, B\. Ding, and J\. Zhou \(2025\)AgentEvolver: Towards Efficient Self\-Evolving Agent System\.arXiv:2511\.10395\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.12.6.1),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.4.3.1)\.
- A\. Zhang, J\. Guo, F\. Jia, C\. Lv, B\. Wang, and F\. Li \(2026a\)EvoAgent: An Evolvable Agent Framework with Skill Learning and Multi\-Agent Delegation\.arXiv:2604\.20133\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.55.49.1)\.
- B\. Zhang, K\. Lazuka, and M\. Murag \(2025\)Equipping Agents for the Real World with Agent Skills\.Note:Anthropic Engineering BlogPublished October 16, 2025; updated December 18, 2025External Links:[Link](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)Cited by:[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§10\.1](https://arxiv.org/html/2607.10113#S10.SS1.p1.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- D\. Zhang, X\. Liu, L\. Cheng, Y\. Wang, K\. Murray, and H\. Wei \(2026b\)SELAUR: Self Evolving LLM Agent via Uncertainty\-aware Rewards\.arXiv:2602\.21158\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.64.58.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.18.18.18.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§9\.2](https://arxiv.org/html/2607.10113#S9.SS2.p1.5)\.
- D\. Zhang \(2026\)AgentDevel: Reframing Self\-Evolving LLM Agents as Release Engineering\.arXiv:2601\.04620\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.87.81.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2)\.
- G\. Zhang, S\. Yang, X\. Gong, and D\. Wang \(2026c\)STARS: Skill\-Triggered Audit for Request\-Conditioned Invocation Safety in Agent Systems\.arXiv:2604\.10286\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.119.113.1),[§11\.1](https://arxiv.org/html/2607.10113#S11.SS1.SSS0.Px5.p1.1),[§12\.7](https://arxiv.org/html/2607.10113#S12.SS7.p1.2)\.
- H\. Zhang, S\. Fan, H\. P\. Zou, Y\. Chen, Z\. Wang, J\. Zhou, C\. Li, W\. Huang, Y\. Yao, K\. Zheng, X\. Liu, X\. Li, and P\. S\. Yu \(2026d\)CoEvoSkills: Self\-Evolving Agent Skills via Co\-Evolutionary Verification\.arXiv:2604\.01687\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.36.30.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p4.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4)\.
- H\. Zhang, Q\. Long, J\. Bao, T\. Feng, W\. Zhang, H\. Yue, and W\. Wang \(2026e\)MemSkill: Learning and Evolving Memory Skills for Self\-Evolving Agents\.arXiv:2602\.02474\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.46.40.1)\.
- M\. Zhang, T\. Shen, H\. Luo, W\. Liu, Z\. Xiao, E\. Cambria, and X\. Tang \(2026f\)SkillFlow: Flow\-Driven Recursive Skill Evolution for Agentic Orchestration\.arXiv:2605\.14089\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.17.11.1),[§9\.2](https://arxiv.org/html/2607.10113#S9.SS2.p1.5)\.
- X\. Zhang, Z\. Liu, Y\. Zhang, X\. Hu, and W\. Shao \(2026g\)RetroAgent: From Solving to Evolving via Retrospective Dual Intrinsic Feedback\.arXiv:2603\.08561\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.20.14.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.30.30.33.3.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§7\.1](https://arxiv.org/html/2607.10113#S7.SS1.p1.1)\.
- X\. Zhang, G\. Wang, Y\. Cui, W\. Qiu, Z\. Li, B\. Zhu, and P\. He \(2026h\)Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents\.arXiv:2604\.15877\.Cited by:[§2\.1](https://arxiv.org/html/2607.10113#S2.SS1.p1.1)\.
- Z\. Zhang, S\. Lu, H\. Qian, D\. He, and Z\. Liu \(2026i\)AgentFactory: A Self\-Evolving Framework Through Executable Subagent Accumulation and Reuse\.arXiv:2603\.18000\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.34.28.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.3.3.3.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1)\.
- Z\. Zhang, K\. Shi, S\. Huang, A\. Nie, Y\. Zeng, Y\. Zhao, Z\. Fang, Q\. Su, H\. Qiu, W\. Yang, Q\. Ren, S\. Zou, W\. Huang, L\. Chen, Z\. Chen, and F\. Zhao \(2026j\)SkillFlow: Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents\.arXiv:2604\.17308\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.83.77.1),[§1\.2](https://arxiv.org/html/2607.10113#S1.SS2.p1.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.13.13.13.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- B\. Zheng, M\. Y\. Fatemi, X\. Jin, Z\. Z\. Wang, A\. Gandhi, Y\. Song, Y\. Gu, J\. Srinivasa, G\. Liu, G\. Neubig, and Y\. Su \(2025a\)SkillWeaver: Web Agents can Self\-Improve by Discovering and Honing Skills\.arXiv:2504\.07079\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.29.23.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§1](https://arxiv.org/html/2607.10113#S1.p5.1),[§9\.1](https://arxiv.org/html/2607.10113#S9.SS1.p1.4),[§9\.4](https://arxiv.org/html/2607.10113#S9.SS4.p1.1)\.
- J\. Zheng, X\. Cai, Q\. Li, D\. Zhang, Z\. Li, Y\. Zhang, L\. Song, and Q\. Ma \(2025b\)LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners\.arXiv:2505\.11942\.Cited by:[§8\.3](https://arxiv.org/html/2607.10113#S8.SS3.p1.1)\.
- J\. Zheng, C\. Shi, X\. Cai, Q\. Li, D\. Zhang, C\. Li, D\. Yu, and Q\. Ma \(2025c\)Lifelong Learning of Large Language Model based Agents: A Roadmap\.arXiv:2501\.07278\.Cited by:[§1\.2](https://arxiv.org/html/2607.10113#S1.SS2.p3.1),[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p1.1)\.
- Y\. Zheng, Z\. Zhang, C\. Ma, Y\. Yu, J\. Zhu, Y\. Wu, T\. Xu, B\. Dong, H\. Zhu, R\. Huang, and G\. Yu \(2026\)SkillRouter: Skill Routing for LLM Agents at Scale\.arXiv:2603\.22455\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.78.72.1),[§1](https://arxiv.org/html/2607.10113#S1.p2.1),[§4](https://arxiv.org/html/2607.10113#S4.p4.3),[§7\.3](https://arxiv.org/html/2607.10113#S7.SS3.p1.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p2.5),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.SSS0.Px1.p1.1),[§9\.3](https://arxiv.org/html/2607.10113#S9.SS3.p1.1)\.
- S\. Zhong, Y\. Lu, J\. Ning, Y\. Wan, L\. Feng, Y\. Ao, L\. F\. R\. Ribeiro, M\. Dreyer, S\. Ammirati, and C\. Xiong \(2026\)SkillLearnBench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real\-World Tasks\.arXiv:2604\.20087\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.96.90.1),[§8\.1](https://arxiv.org/html/2607.10113#S8.SS1.p3.1),[§9\.5](https://arxiv.org/html/2607.10113#S9.SS5.p1.1)\.
- H\. Zhou, S\. Guo, A\. Liu, Z\. Yu, Z\. Gong, B\. Zhao, Z\. Chen, M\. Zhang, Y\. Chen, J\. Li, R\. Yang, Q\. Liu, X\. Yu, J\. Zhou, N\. Wang, C\. Sun, and J\. Wang \(2026a\)Memento\-Skills: Let Agents Design Agents\.arXiv:2603\.18743\.Cited by:[Table 11](https://arxiv.org/html/2607.10113#A1.T11.5.23.17.1),[Table 1](https://arxiv.org/html/2607.10113#S1.T1.10.10.10.2),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p2.1),[§3\.1](https://arxiv.org/html/2607.10113#S3.SS1.p3.1),[§3\.2](https://arxiv.org/html/2607.10113#S3.SS2.SSS0.Px2.p1.2),[Table 5](https://arxiv.org/html/2607.10113#S7.T5.1.1.11.10.1)\.
- Y\. Zhou, S\. Wang, Y\. Su, W\. Du, Y\. Fang, and X\. Lin \(2026b\)A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications\.arXiv:2605\.07358\.Note:Version 1, 8 May 2026Cited by:[§1\.2](https://arxiv.org/html/2607.10113#S1.SS2.p3.1),[§2\.6](https://arxiv.org/html/2607.10113#S2.SS6.p1.1)\.

## Appendix ACoding Protocol and Audit Materials

This appendix records the audit machinery behind the taxonomy\. Method comparisons in the main text are based on mechanisms and evidence roles rather than on recency\.

### A\.1Screening and Coding Fields

Each note was coded for: problem framing, artifact type, update locus, trigger, operator repertoire, learning signal, storage topology, verification or admission mechanism, evaluation setting, headline results, ablations, limitations, closest peers, and survey takeaway\. Papers were included in the primary method cluster if they changed a skill artifact or evaluated the lifecycle of such artifacts\. Modern papers were treated as boundary/context within the broader audit set if they studied generic lifelong learning, generic tool use, memory\-only agents, or adjacent self\-evolution settings without an externally invocable skill artifact\. Older options and hierarchical\-RL papers were cited as background anchors, not coded as modern audit records\.

The coding fields intentionally separate*what changes*from*how evidence supports the synthesis*\. For example, a paper may be coded as an executable\-library method because it edits code skills, while its evidence role may be benchmark, deployment measurement, or safety audit depending on what the paper actually measures\. This separation is why the main text avoids cross\-paper leaderboards unless the benchmark harness is shared\.

### A\.2Evidence Grades

The evidence grades used in Table[7](https://arxiv.org/html/2607.10113#S9.T7)are qualitative audit labels, not statistical confidence intervals\. Grade A indicates multiple controlled ablations, or one clean ablation plus independent corroboration\. Grade B indicates one controlled study, a strong benchmark protocol, or a deployment\-scale measurement\. Grade C indicates convergent benchmark behavior without a clean causal ablation\. Grade D indicates architectural corroboration only\. Mixed labels such as A/B or B/C are used when the supporting papers differ in strength across artifact families\.

### A\.3Master Coding Sheet

TableLABEL:tab:masteris the full representative coding sheet used to audit the lifecycle taxonomy\. It is placed in the appendix because its role is reproducibility rather than narrative: the main text uses the family table and heatmap for synthesis, while this table gives reviewers a way to trace each representative system back to the seven coding fields\.

Table 11:Master coding table for representative dynamic\-skill systems across the seven coding fields of Section[6](https://arxiv.org/html/2607.10113#S6), plus family cluster and one factual headline\.Cells use the legend below; the operator column uses single letters from the ten\-element vocabulary of Equation[4](https://arxiv.org/html/2607.10113#S3.E4)\. The table is a coding sheet that supports the lifecycle\-family taxonomy rather than the primary conceptual taxonomy\.
Legend\.*Artifact:*Code = executable code, NL = natural\-language heuristic, MD = SKILL\.md, LoRA = parametric adapter, Mix = two or more\.*Clock:*fast = in\-context, slow = parametric, 2TS = two\-timescale\.*Trigger:*Task = task\-end retrospective, Fail = failure\-driven, Per = periodic maintenance, User = author/user edit, RL = inside RL loop\.*Operators:*A=Add, R=Refine, M=Merge, S=Split, P=Prune, D=Distill, B=Abstract, C=Compose, W=Rewrite, K=Rerank\.*Signal:*Rew = env reward, Exec = execution feedback, Crit = NL critique, Judge = verifier/judge, XUser = cross\-user aggregation, Teach = teacher distillation\.*Storage:*flat, tree, DAG, graph, subsp, ontol\.MethodClusterArtif\.ClockTrig\.OperatorsSignalStoreHeadlineFoundationalVoyager\(Wanget al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib1)\)Found\.CodefastTaskARew\+Execflat3\.3×\\timesitems; 15\.3×\\timestech\-treeLATM\(Caiet al\.,[2023](https://arxiv.org/html/2607.10113#bib.bib2)\)Found\.CodefastTaskAExecflatTool\-maker \+ tool\-user beats few\-shotSkill\-aware RLSAGE\(Wanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib3)\)RLCode2TSTaskA,RRewflatSkill\-integrated RL on AppWorldSkillRL\(Xiaet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib4)\)RLCodeslowRLA,DRewtreeHier\. skill\-conditioned RLAgentEvolver\(Zhaiet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib5)\)RLMix2TSRLA,R,DRew\+CritflatSelf\-question / navigate / attributeCODE\-SHARP\(Bornemannet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib6)\)RLCode2TSFailA,R,M,WRew\+JudgeDAG4 mutations \+ prereq DAGAgentic\-Prop\.\(Jiaoet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib57)\)RLMixslowRLA,R,P,D,B,CTeach\+JudgeDAG91\.6% AIME\-25 \(30B, 11K traj\.\)COS\-PLAY\(Wuet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib84)\)RLMix2TSRLA,R,P,DRewflatCo\-evolves decision \+ skill bankSkillMaster\(Yanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib117)\)RLMixslowRLA,R,KRew\+ExecflatCounterfactual skill utilitySkillFlow\-Rec\.\(Zhanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib121)\)RLMix2TSRLA,R,P,KRewflatFlow diagnostics for evolutionHeuristic / lesson memoryERL\(Allardet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib7)\)LessonNLfastTaskA,RCritflatTask\-end retrospective lessonsRetroAgent\(Zhanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib8)\)LessonNLfastFailA,RCrit\+JudgeflatDual intrinsic feedbackMUSE\(Yanget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib9)\)LessonNLfastPerA,R,DCrittreeLong\-horizon productivity memoryEvolveR\(Wuet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib11)\)LessonNLfastTaskA,R,WCritflatStrategic\-principle evolutionMemento\(Zhouet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib10)\)LessonMDfastTaskA,R,KCrit\+JudgetreeCase memory \+ SKILL\.md rerankXSkill\(Jianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib12)\)LessonMixfastTaskA,R,M,P,B,KCrit\+JudgeflatVisual dual skill/exp\. storeSAGER\(Taoet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib97)\)LessonNLfastUserA,R,WXUserflatPer\-user policy skills for rec\.EmbodiSkill\(Juet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib118)\)LessonNLfastFailR,WCrit\+ExecflatSkill\-aware embodied reflectionSkillTTA\(Wanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib123)\)LessonNLfastTaskB,D,KExecflatTest\-time transient skillsExecutable skill librariesSkillWeaver\(Zhenget al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib13)\)ExecLibCodefastTaskA,R,PJudge\+ExecflatPropose\-practice\-synth\.\-honeASI\(Wanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib100)\)ExecLibCodefastTaskA,R,B,CExecflatWebArena \+23\.5 vs staticWebXSkill\(Wanget al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib101)\)ExecLibMixfastTaskA,R,D,B,KExec\+JudgegraphGrounded / guided web skillsContractSkill\(Luet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib102)\)ExecLibMixfastFailA,R,BExecflatVWA GLM 28\.1 vs 9\.4 self skillSkillDroid\(Chenet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib103)\)ExecLibCodefastFailA,R,D,B,KExecflat85\.3%, 49% fewer LLM callsAgentFactory\(Zhanget al\.,[2026i](https://arxiv.org/html/2607.10113#bib.bib14)\)ExecLibCodefastUserA,R,CJudgeDAGSubagents as evolvable skillsEvoSkill\(Alzubiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib15)\)ExecLibCodefastFailA,R,PJudge\+ExecflatFailure\-driven \+ Pareto admissionCoEvoSkills\(Zhanget al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib16)\)ExecLibMDfastTaskA,R,BJudgeflatCo\-evolving verifier loopPSN\(Shiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib17)\)ExecLibCodefastFailA,R,M,W,K,PCrit\+Judgegraph5 refactors \+ symbolic creditSkillCraft\(Chenet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib18)\)ExecLibCodefastTaskA,C,RExec\+JudgeDAGVerified compositional MCPLIVE\-SWE\(Xiaet al\.,[2025a](https://arxiv.org/html/2607.10113#bib.bib19)\)ExecLibCodefastTaskA,RExecflatRuntime scaffold synthesisCASCADE\(Huanget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib20)\)LessonMixfastPerA,D,PTeachtreeScientific skill consolidationSkillX\(Wanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib21)\)ExecLibCodefastPerA,S,B,KJudgetreeAutomatic hierarchical libraryTrace2Skill\(Niet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib22)\)ExecLibMDfastTaskA,M,DJudgeflatPrevalence\-weighted consolidationSkillClaw\(Maet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib23)\)ExecLibMDfastUserA,R,M,KXUserflatCross\-user skill evolutionAutoSkill\(Yanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib24)\)ExecLibMDfastUserA,RXUserflatTraining\-free personalized bankAutoRefine\(Qiuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib25)\)ExecLibMDfastPerA,R,MCrit\+JudgeflatDual\-form skills \+ maintenanceMemSkill\(Zhanget al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib26)\)ExecLibMDfastTaskA,R,KCrittreeMeta\-memory skill banksWild\-Skills\(Liuet al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib27)\)ExecLibCodefastPerA,P,KRew\+JudgeflatUtility under realistic retrievalCUA\-Skill\(Chenet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib28)\)ExecLibCodefastTaskA,BExecflatParameterized GUI skillsABSTRAL\(Songet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib29)\)InfraMDfastTaskA,R,CCritgraphMulti\-agent design via SKILL\.mdSkillFoundry\(Shenet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib74)\)ExecLibMixfastPerA,R,M,P,BExec\+JudgetreeScientific contracts \+ provenanceSkillForge\(Liuet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib75)\)ExecLibMDfastFailA,R,WJudgeflatFailure diagnosis in cloud supportSkillMOO\(Gonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib76)\)ExecLibMDfastFailR,P,KJudge\+ExecflatPareto pass/cost optimizationBilevel\-MCTS\(Huanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib90)\)ExecLibMDfastPerR,WJudgeflatMCTS over package structureMetasurface\(Huanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib86)\)ExecLibMixfastFailA,RExecflatPhysics\-verified scientific skillsEvoAgent\(Zhanget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib88)\)ExecLibMixfastUserA,R,K,CXUser\+JudgetreeMulti\-file skills \+ delegationMACRO\(Fanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib99)\)ExecLibCode2TSTaskA,C,DRew\+ExecflatDiscovers composite medical toolsUni\-Skill\(Xieet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib113)\)ExecLibMixfastTaskA,R,D,B,C,KExectreeSkillFolder robotic expansionSkillGen\(Maet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib119)\)ExecLibMDfastTaskA,R,BExec\+JudgeflatNet\-effect verified synthesisHASP\(Liuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib124)\)ExecLibCodefastFailA,R,C,DExec\+TeachflatProgram\-function interventionsSkillOpt\(Yanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib128)\)ExecLibMDfastFailR,WExec\+JudgeflatHeld\-out text\-space optimizerMUSE\-Autoskill\(Linet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib129)\)ExecLibMDfastTaskA,R,KExec\+JudgeflatSkill\-level memory \+ testsSkillGrad\(Wanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib130)\)ExecLibMDfastFailR,WExec\+JudgeflatText gradients \+ momentumParametric / training\-timeSELAUR\(Zhanget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib30)\)ParamLoRAslowFailDRew\(unc\)subspUncertainty\-reshaped fail\-rewardsSKILL0\(Luet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib31)\)ParamLoRAslowPerD,BTeachsubspHelpfulness\-decaying curriculaLSE\(Chenet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib32)\)ParamMixslowTaskD,RRewsubspTrained prompt\-context evolutionSkillsCrafter\(Wanget al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib33)\)ParamLoRAslowPerD,M,KRewsubspLoRA bank \+ semantic subspaceK2\-Agent\(Wuet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib34)\)ParamMix2TSTaskA,R,DRew\+TeachtreeDecl\.\-proc\. co\-evolutionMetaClaw\(Xiaet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib35)\)ParamMix2TSPerA,R,DRewflatFast\-skill / slow\-weight adapt\.Co\-Evolving\(Junget al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib36)\)ParamMixslowFailA,D,PRew\+CritflatHard\-negative failure trainingAgent0\(Xiaet al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib37)\)ParamMix2TSRLA,DRewflatZero\-data curric\.\-executor co\-evTool\-R0\(Acikgozet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib38)\)ParamMixslowRLA,DRewflatDual self\-play for tool\-useSCALAR\(Zabounidiset al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib39)\)ParamMixslowRLD,BTeachflatCo\-adapt difficulty \+ envARISE\(Liet al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib40)\)ParamLoRAslowRLD,KRewsubspSwarm PPO \+ PSO actionsEXIF\(Yanget al\.,[2025b](https://arxiv.org/html/2607.10113#bib.bib41)\)ParamLoRAslowRLD,BRew\+TeachflatExploration\-first skill dataInfrastructure & governanceAgentSkillOS\(Liet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib42)\)InfraMDfastPerA,C,S,P,KJudgetreeCapability tree \+ DAG orch\.SkillRouter\(Zhenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib43)\)InfraCodefastPerKJudgeflatFull\-text retrieval 80K skillsSkVM\(Chenet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib44)\)InfraCodefastUserA,R,CExecDAGSkills as compilable codeSkillNet\(Lianget al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib45)\)InfraMDfastPerA,K,PJudgegraphOntology \+ relation graphSkillOrchestra\(Wanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib46)\)InfraMDfastPerK,CJudgeontolSkill handbooks for routingSkillFlow\-2025\(Liet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib47)\)InfraMDfastTaskKJudgeflatMulti\-stage community\-skill retrievalSkillFlow\-Bench\(Zhanget al\.,[2026j](https://arxiv.org/html/2607.10113#bib.bib48)\)EvalMDfastTaskA,R,PCrit\+Judgeflat166 tasks / 20 task familiesSWE\-Skills\-Bench\(Hanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib104)\)EvalMDfastTaskKExecflat49 SWE skills; \+1\.2 avg\. gainSRA\(Suet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib111)\)EvalMDfastTaskKExec\+Judgeflat5\.4K tasks / 26K\-skill corpusAutoAgent\(Wanget al\.,[2026e](https://arxiv.org/html/2607.10113#bib.bib49)\)InfraMix2TSTaskA,R,CCrit\+JudgeDAGEvolves tools, peers, and selfAgentDevel\(Zhang,[2026](https://arxiv.org/html/2607.10113#bib.bib50)\)InfraMDfastUserA,R,K,PXUsergraphRelease engineering \+ lineageGEA\(Wenget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib51)\)InfraMix2TSPerA,R,MJudge\+RewontolGroup\-based framework evolutionSingle\-Ag\.Sk\.\(Li,[2026](https://arxiv.org/html/2607.10113#bib.bib52)\)InfraCodefastPerA,M,D,PJudgeflatMulti→\\tosingle compilationEffiSkill\(Wanget al\.,[2026i](https://arxiv.org/html/2607.10113#bib.bib53)\)InfraCodefastPerA,B,KRewflatMined operator skillsGraph of Skills\(Liuet al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib89)\)InfraCodefastUserK,CRewgraphDependency\-aware retrievalGraSP\(Xiaet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib83)\)InfraCodefastTaskC,RExecDAGTyped DAG composition \+ repairSkilldex\(Saha and Hemanth,[2026](https://arxiv.org/html/2607.10113#bib.bib81)\)InfraMDfastUserA,K,PJudgetreePackage manager \+ scoped registryCorpus2Skill\(Sunet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib79)\)InfraMDfastUserA,S,KJudgetreeNavigable enterprise QA skillsSkillRepoMining\(Biet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib96)\)InfraMDfastUserA,BJudgeflatGitHub repo mining to SKILL\.mdAgentSkills\-Data\(Linget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib87)\)EvalMDfastPerK,PJudge—40K\+ public\-skill ecosystem studySSL\(Lianget al\.,[2026a](https://arxiv.org/html/2607.10113#bib.bib112)\)InfraMDfastUserB,KJudgegraphMRR \.649→\\to\.729; F1 \.409→\\to\.509SkillLearnBench\(Zhonget al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib82)\)EvalMDfastTaskA,RJudgeflatContinual skill\-generation benchmarkRaw\-Experience\(Huanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib127)\)EvalMDfastTaskB,DExec\+JudgeflatExtraction\-consumption lifecycle studySkillsVote\(Liuet al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib125)\)InfraMixfastTaskA,R,P,KExec\+JudgeflatGoverned open\-skill evolutionSkillSmith\(Xuet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib122)\)InfraMDfastUserB,C,KExectreeBoundary\-guided runtime interfaceSkillOps\(Songet al\.,[2026b](https://arxiv.org/html/2607.10113#bib.bib132)\)InfraCodefastPerR,M,P,K,CExec\+JudgegraphTechnical\-debt maintenance layerSkillsInjector\(Liet al\.,[2026d](https://arxiv.org/html/2607.10113#bib.bib131)\)InfraMDfastTaskK,WExecflatAdaptive skill\-context constructionSafety & auditASG\-SI\(Huang and Huang,[2025](https://arxiv.org/html/2607.10113#bib.bib55)\)SafetyMix2TSTaskA,R,P,DRew\+JudgegraphAudited skill graph \+ verif\. rewardsClawSafety\(Weiet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib56)\)SafetyMDfastUserPJudge—Highest\-ASR vector in personal agentsSecure\-Skills\(Liet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib72)\)SafetyMDfastUserPJudge—Lifecycle threat taxonomyAgentSkills\-PI\(Schmotzet al\.,[2025](https://arxiv.org/html/2607.10113#bib.bib105)\)SafetyMDfastUserPExec—Simple SKILL\.md injection channelSupply\-Chain\(Quet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib73)\)SafetyMDfastUserPJudge—DDIPE poisoning attacksAgentSkills\-Wild\(Liuet al\.,[2026g](https://arxiv.org/html/2607.10113#bib.bib106)\)SafetyMDfastUserPJudge—31K skills; 26\.1% vuln\.Malicious\-Skills\-Wild\(Liuet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib107)\)SafetyMixfastUserPExec\+Judge—157 confirmed malicious skillsSkill\-Inject\(Schmotzet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib108)\)SafetyMixfastUserPJudge—202 injection\-task pairsSkillJect\(Jiaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib109)\)SafetyMDfastUserR,W,PExec\+Judge—95\.1% automated ASRSkillSieve\(Hou and Yang,[2026](https://arxiv.org/html/2607.10113#bib.bib77)\)SafetyMDfastUserPJudge—Hierarchical malicious\-skill triageSkillAttack\(Duanet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib78)\)SafetyMDfastUserPJudge—Automated attack\-path red teamingBadSkill\(Tieet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib91)\)SafetyMixfastUserPExec—Model\-in\-skill backdoorsCredLeak\(Chenet al\.,[2026f](https://arxiv.org/html/2607.10113#bib.bib92)\)SafetyMixfastUserPExec\+Judge—Cross\-modal secret leakageHarmfulSkillBench\(Jianget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib80)\)SafetyMDfastUserPJudge—Harmful\-but\-valid skillsSkillStealing\(Wanget al\.,[2026h](https://arxiv.org/html/2607.10113#bib.bib85)\)SafetyMDfastUserKJudge—Black\-box skill extractionMalicious\-or\-Not\(Holzbaueret al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib95)\)SafetyMDfastUserPJudge—Repository\-aware classificationSTARS\(Zhanget al\.,[2026c](https://arxiv.org/html/2607.10113#bib.bib110)\)SafetyMDfastUserK,PJudge—Request\-conditioned invocation auditMedSkillAudit\(Houet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib93)\)SafetyMDfastUserP,RJudge—Medical release\-readiness auditSemantic\-Supply\(Sahaet al\.,[2026](https://arxiv.org/html/2607.10113#bib.bib120)\)SafetyMDfastUserK,P,WJudge—SKILL\.md semantic registry attacksUser\-Comprehension\(Wen,[2026](https://arxiv.org/html/2607.10113#bib.bib126)\)SafetyMDfastUserRJudge—Skill\-spec disclosure audit