The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents

# Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents
Source: [https://arxiv.org/html/2608.06663](https://arxiv.org/html/2608.06663)
Mingguang Chen1,∗\*Licheng Wang2Bo Qu1 1DeepGrounding2AlphaAvatar ∗\*Corresponding author\. Email:[deepgroundingai@gmail\.com](https://arxiv.org/html/2608.06663v1/mailto:[email protected])

\(July 2026\)

###### Abstract

Frontier language models solve, in a single forward pass, reasoning problems that would have been research contributions a few years ago — yet the same models, embedded in an agent loop and asked to complete a task spanning hours rather than seconds, fail in ways no single\-step benchmark reveals: losing track of an earlier decision, declaring a half\-finished job done, or quietly drifting from the goal they were given\. We call the distance between single\-step capability and reliable long\-task completion thehorizon gap, and survey1,547 arXiv papers \(2024\-2026\)— collected via a systematic eight\-thread seed harvest with a disclosed, quantified two\-stage bleed filter \(26\.8% of raw hits excluded as off\-topic\), extended by a targeted supplement into a badly under\-covered theory/safety category — that map the field’s response to it\. We first disambiguate three properties the literature routinely conflates:*long\-horizon*\(a property of the task, its required number of steps\),*long\-context*\(a property of the model, how many tokens it can attend to at once\), and*long\-term memory*\(a property of the system, whether information persists across steps or sessions\) — logically independent axes that a single “long\-horizon” label obscures\. We then organize the corpus into six categories that track a long\-horizon task’s lifecycle: planning and decomposition, memory and context management, execution control and recovery, training for long horizons, evaluation and measurement, and the foundations, limits, and safety of running agents unattended for extended periods — crossed with a second axis,*where the horizon is carried*\(within one context, within one task via a harness that exceeds the context window, or persistently across tasks and sessions\), that organizes every technical section\. Across all six, we find the same structural pattern: outcome\-only signals — a single reward, a single pass/fail check — grow uninformative as horizon grows, and the field’s response, whether in training \(process reward models and credit assignment\) or evaluation \(trajectory\-level diagnostics that supersede pass/fail benchmarks\), is to manufacture denser, step\-level signal in its place\. We treat the critical and diagnostic literature — demonstrations that self\-correction, benchmark scores, or training signals do not mean what they are assumed to mean — as a first\-class thread throughout rather than a separate critique chapter, since a survey that segregated critique from method in this field would routinely split single papers across two chapters\. We close by naming what we take to be the field’s most consequential open measurement problems: how much long\-horizon capability lives in the underlying model versus the harness wrapped around it, the risk of correlated measurement bias where the process\-level signals used to train long\-horizon agents and those used to evaluate them rest on shared assumptions about what counts as progress, and whether long\-horizon reliability admits any general predictive theory at all\.

## 1\. Introduction

A frontier language model can now solve, in a single forward pass, reasoning problems that would have been research contributions a few years ago\. The same model, dropped into an agent loop and asked to keep a multi\-hour software\-engineering task on track, still fails in ways a junior engineer would not — losing track of an earlier decision, declaring victory on a half\-finished patch, or quietly drifting from the goal it was given\. Measurement work tracking the length of task \(in human\-time\-to\-complete\) that frontier models can complete at a fixed reliability threshold reports this length growing at a measured, roughly exponential rate across recent model generations\[[45](https://arxiv.org/html/2608.06663#bib.bib1571)\]— a striking trend we treat, per this survey’s own language calibration \(§2\.2\), as*measured and reported*, not as a settled law, since what counts as “the same” task across model eras and whether human\-time\-to\-complete on a benchmark generalizes to messier real deployments are both contested\. But even taking the trend at face value, it describes model capability outpacing something else: the reliable*completion*of long tasks in deployment, where harness failures, evaluation gaps, and accumulating error still bite in ways single\-step benchmarks never see\. We call this thehorizon gap— the distance between what a model can do in one step and what a system built around it can reliably finish over many — and argue it is the primary bottleneck standing between current agent capability and durable real\-world deployment\.

This survey maps the research literature responding to that gap\. We assembled a corpus of1,547 arXiv papers \(2024\-2026\)— a systematic seed harvest across eight threads of the long\-horizon\-agent literature \(1,419 papers, after a two\-stage bleed filter removed 26\.8% of raw hits judged off\-topic, primarily classical multi\-agent\-RL, robotic manipulation, and time\-series\-forecasting work this survey explicitly scopes out\), extended by a targeted supplemental harvest of 128 papers aimed at directions the reclassification exposed as under\-covered, chiefly the theory and safety literature of §8 — and classified all of them intosix technical categories\(Table 1, Figure 1\)\. Three observations motivate the paper’s structure\.

First, the field’s vocabulary conflates three logically independent properties\.“Long\- horizon,” “long\-context,” and “long\-term memory” are used almost interchangeably in casual discussion, but a task’s horizon \(how many steps it requires\), a model’s context length \(how many tokens it can attend to at once\), and a system’s memory \(whether information persists across steps or sessions\) can vary independently — a harness can carry a long\-horizon task through a short context via aggressive summarization, and a system can have a huge context window and no persistent memory at all\. §2 fixes this and several other overloaded terms \(agent vs\. harness, episode vs\. session, coherence, autonomy time\) before using them, because imprecision here is not merely stylistic: it is the reason a “memory” paper and a “long\-context” paper are so often cited as if they addressed the same problem when they do not\.

Second, no single component is where long\-horizon tasks succeed or fail — the whole pipeline is\.We organize the survey around six categories that track a task’s lifecycle:*planning*decides what to do \(§3\);*memory*decides what information that decision draws on \(§4\);*execution control*decides how the resulting actions are run and recovered from when they fail \(§5\);*training*decides how the underlying policy learns to act well over many steps \(§6\);*evaluation*decides whether any of this actually works \(§7\); and*foundations*covers the theory of why long horizons degrade performance at all, and the oversight problem of running an agent for a long time without a human watching every step \(§8\)\. Existing surveys cover individual pieces of this pipeline — planning strategies, memory mechanisms, GUI\-agent methods, agentic reinforcement learning — in isolation \(§2\.4\); this survey’s contribution is to organize the full pipeline around one cross\-cutting question asked of every stage:*what breaks first as the required horizon grows, and what compensates?*

Third, the critical and diagnostic literature is not a footnote — it is often the most important literature in each category\.A recurring pattern across §§3\-8 is that the most consequential recent papers are not new methods but rigorous demonstrations that an earlier method, benchmark, or metric does not do what it claims: intrinsic self\-correction does not reliably improve reasoning without external grounding\[[31](https://arxiv.org/html/2608.06663#bib.bib1573)\]; a substantial share of reported SWE\-bench solves reflect test\-suite weakness or training\-data leakage rather than genuine issue resolution\[[94](https://arxiv.org/html/2608.06663#bib.bib240),[56](https://arxiv.org/html/2608.06663#bib.bib354)\]; and agentic benchmarks broadly have systematic construction problems that best practices are only now starting to name explicitly\[[136](https://arxiv.org/html/2608.06663#bib.bib382)\]\. We treat this diagnostic thread as a first\-class citizen throughout rather than segregating it into a single critique section, because in this literature the critique of a method and the state of the art in that method’s category are often the same papers\.

Contributions\.\(1\) A disambiguation of long\-horizon, long\-context, and long\-term memory, and a six\-category taxonomy — planning, memory, execution, training, evaluation, foundations — crossed with a second axis \(where the horizon is carried: within\-context, within\-task\-beyond\- context, or cross\-task\-persistent\) that organizes every technical section \(§2\)\. \(2\) A systematic map of 1,547 papers into that taxonomy, with a disclosed, quantified query\-bleed rate and an explicit two\-stage construction \(§2\.2, Figure 2, Figure 6\)\. \(3\) A synthesis identifying two specific unresolved measurement problems — how much of long\-horizon capability lives in the model versus the harness \(§5, §9\), and the risk of correlated measurement bias between the process\-level signals used to train long\-horizon agents and those used to evaluate them \(§6, §7, §9\) — together with a cross\-cutting observation that the execution*trajectory*, not the final outcome, is becoming the field’s shared unit of analysis \(§9\)\. \(4\) An explicit positioning of this survey against the recursive\-self\-improvement literature it borders but does not overlap with, self\-cited and scoped in §2\.4 and §6\.

## 2\. Preliminaries

### 2\.1 Definitions

The literature on long\-horizon agents uses a small set of terms so loosely that surveys risk adding noise rather than removing it\. We fix working definitions here and use them consistently for the rest of the paper\.

Ataskis a specification of a desired end state together with a checkable completion criterion — “resolve this GitHub issue,” “book an itinerary satisfying these constraints,” “reduce this eval’s error rate\.” A task’shorizonis the number of sequential decisions, actions, or environment interactions required to go from the initial state to task completion; it is a property of the task \(and the environment it is embedded in\), not of any particular system attempting it\. Anepisodeis one complete attempt at a task, from a defined start state to a terminal state — success, failure, or time\-out — the unit of account inherited from reinforcement learning\. Asessionis a continuous interactive period between a user \(or an orchestrating process\) and an agent, bounded by the interface layer rather than by the task: a session may contain zero, one, or many episodes, and a single episode may span multiple sessions if the system’s state persists across the boundary\. Conflating episode and session is a common source of confusion in the memory literature \(§4\), where “long\-term” sometimes means “beyond this episode” and sometimes means “beyond this session\.”

We also distinguish anagent— the policy that decides which action to take next, i\.e\. the model together with whatever prompting/decoding strategy turns its outputs into actions — from itsharness\(also called scaffold\): the surrounding software that converts a single\-forward\-pass model into a system capable of taking many actions over time — the control loop, tool\-execution sandbox, memory read/write layer, retry and recovery logic, and any sub\-agent orchestration\. This distinction matters because a recurring finding across §§3–5 is that harness engineering, not the underlying model, is often the binding constraint on how long a task an agent can complete — the same model dropped into a better harness completes measurably longer tasks\.

The paper’s central disambiguation is among three axes that the field’s vocabulary routinely conflates:

- •Long\-contextis a property of the model/serving stack: how many tokens a single forward pass can attend to\.
- •Long\-horizonis a property of the*task*: how many sequential decisions or actions it requires, independent of whether the resulting interaction trace fits inside one context window\.
- •Long\-term memoryis a property of the*system*: whether information available at step or sessionttremains available \(in some form — verbatim in context, retrieved from an external store, or baked into weights\) at step or sessiont\+kt\+k\.

These are logically independent\. A task can be long\-horizon but short\-context, if a harness aggressively summarizes or discards intermediate state between steps \(at some risk to correctness, §4\)\. A system can have a very long context window and still have no long\-term memory in any interesting sense, if every session starts from a blank slate\. And a system can exhibit long\-term memory without ever facing a long\-horizon task, if it merely caches user preferences across otherwise\-independent short interactions\. Table 1 and Figure 1 organize the survey along a related but distinct axis — where the extra horizon is*carried*— which we introduce in §2\.3\.

Two further terms recur across sections\.Coherencedenotes the property that an agent’s actions across steps remain consistent with a single evolving representation of the goal and the plan toward it, rather than silently drifting, forgetting earlier commitments, or contradicting itself; loss of coherence is the qualitative failure mode that long\-horizon degradation curves are trying to measure quantitatively \(§8\)\.Autonomy time\(or*time horizon*, in the sense popularized by measurement work discussed in §7 and §8\) is an empirical capability metric: the length of task, measured in human\-time\-to\-complete, that an agent can complete at a fixed reliability threshold \(e\.g\., 50%\) — a way of putting a number on “how long a horizon can this system actually carry,” as opposed to the task\-side horizon defined above\.

### 2\.2 Corpus construction

The corpus was built in two stages: a systematic seed harvest, then a targeted supplemental harvest aimed at categories the first stage under\-covered\. Both stages, and the classification step between them, were carried out by a single annotator \(the author\); all classification rules are released as scripts alongside the corpus \(§ Data Availability\), and the noise this process introduces is quantified below rather than assumed away\.

Seed harvest\.We queried the arXiv API across eight threads covering the literature’s recognized sub\-directions — long\-horizon tasks and agents; LLM agent planning and task decomposition; agent memory and context management; agentic reinforcement learning and credit assignment; agent benchmarks and evaluation; multi\-agent orchestration and scaffolding; error accumulation and reliability; and self\-correction and recovery — each restricted tocs\.AI, cs\.CL, cs\.LG, cs\.SE, cs\.MA, cs\.HC, stat\.MLand a submission window of 2024\-01 through 2026\-07 \(2023 foundational work is handled separately as hand\-verified anchors, §2\.3, §5\)\. Each thread was capped at 250 results \(a disclosed depth limit\) and sorted by relevance\. Across the eight threads this returned 2,000 raw hits, 1,939 unique after cross\-thread deduplication by arXiv ID — the low duplicate rate \(61 hits\) indicates the threads are reasonably distinct rather than redundant\. A harvest\-time relevance filter then required title\+abstract to contain at least one topic\-signal term \(e\.g\.*long\-horizon*,*multi\-step*,*episod\-*,*trajector\-*,*planning*,*memory*, a named benchmark,*credit assignment*,*reflection*,*replan\-*,*orchestrat\-*,*scaffold\-*\) so that the single most overloaded word in this literature — “agent” — could not by itself admit a paper; this dropped 176 hits \(9\.1% of the unique pool\), leaving 1,763 seed candidates\.

Query bleed and reclassification\.Because “agent” and several of the thread queries \(especially*credit assignment*and*reinforcement learning for agents*\) also match a large classical literature — multi\-agent reinforcement learning theory, robotic manipulation and motion planning, autoregressive time\-series/PDE forecasting — that this survey explicitly scopes out \(classical HRL, options, and robotic task\-and\-motion planning are treated as historical anchors only, not as corpus members, following the boundary set in §1\), we ran an exploratory TF\-IDF \+ truncated\-SVD \+kk\-means clustering \(k=18k=18\) over the seed pool purely as scaffolding for taxonomy design, then reclassified every paper with cluster\-based category defaults, cross\-cutting keyword rules, and an explicit LLM/agent\-signal gate that catches classical\-literature bleed that the harvest\-time filter’s broader terms let through \(clusters that were agent\-systems\-specific by construction — e\.g\. a GUI\-agent cluster, an agent\-memory\-systems cluster — were exempted from this gate after we verified by inspection that gating them produced false exclusions\)\. This second\-stage filter excluded 344 further papers \(19\.5% of the 1,763 seed candidates\); combined with the harvest\-time filter, 520 of the 1,939 raw unique hits \(26\.8%\) were judged off\-topic and dropped over the two stages — we report this as an absolute number, following the recommendation that corpora built around a term as overloaded as “agent” disclose bleed quantitatively rather than assert its absence\. A random sample of 30 kept papers and 55 excluded papers was manually re\-checked against this criterion; the estimated residual misclassification rate after both filters is on the order of one in twenty, concentrated at the boundary between LLM\-driven and classical robotic/multi\-agent\-RL systems, which is inherently fuzzy for a handful of transitional papers\. This left1,419 seed papers\.

Targeted supplement\.Reclassification exposed one severe gap —foundations\(failure\-mode theory, reliability/scaling, oversight of long\-running agents\) held only 10 papers, well under the threshold at which we would have folded it back into prose\-only treatment inside another section — and two thinner spots \(cross\-session memory persistence; credit\-assignment theory\)\. We ran 19 targeted arXiv queries against these directions \(2024\+,cs\.\*/stat\.ML\), yielding 321 raw candidates; a stricter secondary filter requiring both an explicit agent/agents token and a foundations\-or\-long\-horizon signal term dropped 193 \(60\.1% — a markedly higher bleed rate than the seed harvest, consistent with these being broader, less\-targeted query terms such as “reward hacking” or “scaling law” that catch unrelated work on quantization, pretraining, or physical\-system forecasting\), keeping128 supplement papers, taggedsource=supplement\. The supplement is recency\-biased by construction and is therefore excluded from the growth\-timeline figure \(Figure 6, §9\), which uses seed\-only counts\.

Final corpus and enrichment\.The combined corpus is1,547 papers\(1,419 seed \+ 128 supplement\) — larger than our initial ~1,100–1,400 planning estimate; rather than force the corpus down to a pre\-registered size, we kept the full bleed\-filtered yield\. Author and venue metadata were resolved via batched OpenAlex lookups \(by the arXiv DOI form\), with a mandatory arXiv\-API fallback for the 180 rows \(11\.6%\) OpenAlex had not yet indexed or indexed without authors; this achieved 100% author coverage\. Because the corpus is overwhelmingly recent \(58–77% of most categories was posted in 2026 alone, Table 1\), citation counts would be near zero for the majority of entries and are not a meaningful signal here; we do not report or use them, and caution against reading corpus composition as a proxy for impact anywhere in this paper\.

Single\-label assignment is a simplification\.Each paper carries exactly onecategoryin the released corpus, but a large fraction of this literature is genuinely multi\-topic: a memory\-augmented agent evaluated on a new benchmark and trained with process rewards has a legitimate claim to three categories\. Our labels record the*primary*contribution — the question the paper’s own abstract foregrounds — not an exclusive membership claim\. As a lower\-bound indicator of how often that primary\-contribution call is a real judgment rather than an obvious single fit, we counted how many of the six categories’ topic\-keyword signals each paper’s*title alone*triggers:13\.8% of the corpus \(213 papers\) trips two or more\. We report this as a keyword co\-occurrence statistic, not a semantic scope estimate — it is a lexical measurement, not a claim about what a domain expert would conclude, and it is itself an undercount in one direction \(a title like “X: A Benchmark for Y” trips bothevaluationand Y’s signal even when the paper’s unambiguous primary contribution is the benchmark\) and an overcount in another \(co\-occurring keywords do not always mean co\-equal contributions\)\. We do not extend this count to title\-plus\-abstract, where generic phrases like “we evaluate” fire theevaluationsignal in nearly every machine\-learning abstract regardless of topic, making the resulting number uninformative rather than merely noisy\. The most common title\-level overlaps are memory\+evaluation \(29 papers, chiefly memory benchmarks\), memory\+execution \(23\), and training\+execution \(22\), which is why §§4\-7 cross\-reference each other as often as they do\. Readers should treat per\-category counts as a map of where each paper’s*center of mass*sits, not as disjoint bins, and the release includes full abstracts so that any alternative labeling can be recomputed from the same corpus\.

Limitations\.This is a single\-annotator, largely rule\-based corpus: classification followed cluster defaults, keyword rules, and an explicit per\-paper override list \(never hand\-edits to the underlying data\), all released as scripts\. It under\-represents unpublished industrial agent engineering that never reaches arXiv, non\-English\-language work, and any system described only in a blog post or technical report rather than a paper\. Benchmark leaderboard positions move on a timescale of months; every specific benchmark number we cite is dated at first mention rather than presented as a current ranking\.

Language calibration\.Throughout this paper we avoid unmeasured strong language — “proves,” “monotone,” “law,” “first” — in favor of “observed,” “measured and reported,” or “tracks,” reserving stronger language only for claims a cited paper’s own experiments establish directly\. This applies with particular force to the METR time\-horizon trend \(§7, §9\): we treat its reported growth rate as*measured and reported*, not as a settled law, because the metric’s cross\-model\-era comparability and its generalization beyond benchmark tasks are both contested rather than resolved\.

### 2\.3 Taxonomy

We organize the corpus along one primary axis realized as six categories, plus a second, prose\-level axis that cuts across all six and is this paper’s main organizing device\.

Axis 1 — what the paper is about, thecategorycolumn reported in Table 1:*planning & decomposition*\(§3: how agents turn a task into a sequence of committable steps —decompose/search/worldmodelsubcategories\),*memory & context management*\(§4: how information persists within and across a task —context/externalsubcategories, plus a weights tier discussed qualitatively, not separately tagged\),*execution control & recovery*\(§5: the runtime loop that turns plans into actions and recovers from failures —loop/orchestration/recoverysubcategories\),*training for long horizons*\(§6: how models are trained to act well over many steps —rl/supervisionsubcategories\),*evaluation & measurement*\(§7: how we know whether any of the above works — currently a singlebenchmarksubcategory, since every classified paper in it is benchmark work, plus the critical literature that interrogates benchmarks throughout\), and*foundations, limits & safety*\(§8: theory of degradation and compounding error, and the oversight problem for agents that run for a long time unsupervised — not subdivided into subcategories\)\. Table 1 reports per\-category counts, the subcategory tags each category actually carries, and the fraction of each category posted in 2026 alone\.

Table 1\.Corpus composition by category \(1,547 papers total\)\.

CategorySubcategory tags \(n\)PapersOf which supplement% posted 2026Planning & decompositiondecompose \(162\) · worldmodel \(11\) · search \(9\)182027%Memory & context managementexternal \(294\) · context \(103\)3972172%Execution control & recoveryorchestration \(338\) · recovery \(245\) · loop \(1\)584045%Training for long horizonsrl \(130\) · supervision \(37\)1671462%Evaluation & measurementbenchmark \(114\)114058%Foundations, limits & safetynot subdivided1039377%Subcategory tags are a corpus column for five of six categories;evaluationcurrently resolves to a single tag \(benchmark\) because every classified paper in it is benchmark work, andfoundationswas not given a subcategory split \(§8’s internal organization — compounding error, goal drift/misalignment, oversight — is a prose\-level grouping, not a tagged column, matching how Axis 2 below is handled\)\.

![Figure 1: Taxonomy grid -- six categories (rows) crossed with the horizon-bearing-locus axis (columns). Each cell names representative systems drawn from the corpus and discussed in the corresponding section. The foundations/within-context cell is empty by construction: degradation and oversight phenomena presuppose a task that already exceeds one forward pass.](https://arxiv.org/html/2608.06663v1/figures/figure1_taxonomy_grid.png)Figure 1: Taxonomy grid – six categories \(rows\) crossed with the horizon\-bearing\-locus axis \(columns\)\. Each cell names representative systems drawn from the corpus and discussed in the corresponding section\. The foundations/within\-context cell is empty by construction: degradation and oversight phenomena presuppose a task that already exceeds one forward pass\.Figure 1 renders Axis 1 crossed with Axis 2 as a grid, with representative systems per cell drawn from the corpus and each discussed in the corresponding section below\. Seventeen of the eighteen cells are populated: only thefoundations/within\-context cell is empty, because degradation and oversight problems are, definitionally, phenomena of tasks that already exceed a single forward pass\. The cross\-task\-persistent column is the sparsest in the literature but is populated in every other row, which is worth noting explicitly — the impulse to let capability accumulate across episodes is not confined to the memory literature where it is most visible, but appears as a distinct minority thread inside planning \(§3\), execution \(§5\), training \(§6\), evaluation \(§7\), and the self\-evolution work discussed in §8\. Figure 2 shows the same six categories as a semantic map \(TF\-IDF projection with per\-category density contours\), which makes visible thatexecutionandplanninginterpenetrate heavily — consistent with the two sharing much of their harness\-engineering literature — whiletrainingforms the most topically separated cluster\.

![Figure 2: Semantic map of the corpus (TF-IDF + SVD + t-SNE projection, axes are arbitrary embedding dimensions) with per-category density contours and direct labels. n=1,547.](https://arxiv.org/html/2608.06663v1/figures/map_landscape.png)Figure 2: Semantic map of the corpus \(TF\-IDF \+ SVD \+ t\-SNE projection, axes are arbitrary embedding dimensions\) with per\-category density contours and direct labels\. n=1,547\.Axis 2 — where the horizon is carried, a prose\-level distinction we apply within every technical section rather than as a corpus column:within\-context\(the task’s full history fits in one forward pass — “long\-horizon” here means many steps, not literally beyond\-context\),within\-task\-beyond\-context\(the step count exceeds one context window, but a harness — external memory, summarization, sub\-agent handoff — keeps the attempt inside one continuous episode\), andcross\-task\-persistent\(information or skill survives across distinct episodes or sessions — memory banks, fine\-tuned weights, accumulated skill libraries — so horizon is extended by accumulation over time rather than within a single attempt\)\. Every technical section below is, in part, organized by where along this axis its methods sit, because the engineering problem — and the failure modes — differ sharply across the three: within\-context methods fail by attention/utilization degradation \(§4, §8\); within\-task\-beyond\-context methods fail by information loss at the harness boundary; and cross\-task\-persistent methods fail by interference or staleness in what was retained\.

### 2\.4 Positioning vs\. existing surveys

Several recent surveys cover pieces of this territory\. Planning surveys such as\[[33](https://arxiv.org/html/2608.06663#bib.bib8)\]and\[[90](https://arxiv.org/html/2608.06663#bib.bib244)\]map the space of LLM planning strategies; memory surveys such as\[[130](https://arxiv.org/html/2608.06663#bib.bib25)\]and its 2026 update\[[32](https://arxiv.org/html/2608.06663#bib.bib713)\]chart agent memory mechanisms in detail; GUI/web\-agent surveys such as\[[121](https://arxiv.org/html/2608.06663#bib.bib134)\],\[[70](https://arxiv.org/html/2608.06663#bib.bib146)\], and\[[89](https://arxiv.org/html/2608.06663#bib.bib247)\]cover a single embodiment of the execution problem; and an agentic\-reinforcement\-learning survey\[[122](https://arxiv.org/html/2608.06663#bib.bib457)\]covers one training paradigm in §6’s territory\. Each is organized around a single component of the pipeline\. This survey’s delta is to organize instead aroundhorizonas the cross\-cutting axis: rather than asking “what memory mechanisms exist” or “what planning strategies exist” in isolation, we ask, for every stage of the pipeline,*what breaks first as the required horizon grows, and what compensates*— which is why planning, memory, execution, training, evaluation, and foundations are treated as one continuous argument \(§§3–8\) rather than as independent surveys stitched together, and why the critical/diagnostic literature \(benchmark critiques, failure\-mode analyses, self\-correction skepticism such as\[[38](https://arxiv.org/html/2608.06663#bib.bib39)\]\) is carried as a first\-class thread through every section rather than segregated into its own critique chapter\.

This survey is also a deliberate companion to a prior survey of recursive self\-improvement in AI\[[13](https://arxiv.org/html/2608.06663#bib.bib1574)\], with which it shares harvest infrastructure and taxonomy methodology but not scope: that survey covers loops in which a system improves itself — its outputs, its training data, its evaluator, or the research process — as an end in itself; this survey covers training and engineering aimed specifically at extending how long a task a system can carry out, and treats self\-improvement loops only where they are instrumental to that end \(§6\)\. Readers interested in*why*models are trained on their own generated data, rather than*how far a single trained policy can be pushed on long tasks*, should consult the companion survey; we flag the boundary again where it becomes concrete, in §6\.

## 3\. Planning & Decomposition

Planning is the first place a long\-horizon task must be made tractable: before any action is taken, something has to turn “resolve this issue” or “book this itinerary” into a sequence of committable steps\. The literature spans a spectrum from committing to a full plan before any execution, through interleaving one planning step with one execution step, to explicit search over multiple candidate continuations — and, as the corpus grew, it became clear that where a system sits on this spectrum tracks how uncertain its environment is, not how sophisticated its model is\. Plan\-first decomposition is data\-efficient and interpretable precisely when the environment is predictable enough that a plan made att=0t=0still holds att=kt=k; as uncertainty grows, that assumption breaks, and the field’s center of mass has moved toward interleaved and search\-based alternatives \(this section,decomposen=162n=162,searchn=9n=9,worldmodeln=11n=11of the taxonomy’splanningcategory\)\.

![Figure 3: The planning strategy spectrum -- plan-then-execute, interleaved, search-based, and world-model-based lookahead approaches, positioned against increasing environment uncertainty and action-space size.](https://arxiv.org/html/2608.06663v1/figures/figure3_planning_spectrum.png)Figure 3: The planning strategy spectrum – plan\-then\-execute, interleaved, search\-based, and world\-model\-based lookahead approaches, positioned against increasing environment uncertainty and action\-space size\.Plan\-then\-execute and its failure mode\.The idea that a long task should be attacked by decomposing it into temporally extended sub\-behaviors long predates language models: the options framework formalized temporal abstraction over sub\-policies in reinforcement learning\[[88](https://arxiv.org/html/2608.06663#bib.bib1561)\], and feudal reinforcement learning organized control into a manager hierarchy assigning sub\-goals to workers\[[18](https://arxiv.org/html/2608.06663#bib.bib1562)\]— the structural ancestor of the planner/executor splits that recur throughout this section\. Two 2022\-2023 systems carried the pattern into the LLM setting from opposite directions: LLM\+P routes planning through a classical symbolic planner, using the language model to translate the problem into a formal specification rather than to plan directly\[[59](https://arxiv.org/html/2608.06663#bib.bib1563)\], while SayCan grounds an LLM’s proposed steps in what a robot can actually execute by weighting them against learned affordances\[[4](https://arxiv.org/html/2608.06663#bib.bib1564)\]— an early statement of this section’s recurring lesson that a plan is only as good as its feasibility check\. The classical recipe — decompose the task into sub\-goals, then execute each — remains the default in tool\-use and code\-generation settings where the action space is well\-specified: PoTable applies staged plan\-then\-execute reasoning to table QA\[[64](https://arxiv.org/html/2608.06663#bib.bib136)\], and Tree\-of\-Code consolidates an LLM agent’s actions into a unified code\-based action space precisely to make each planned step verifiable before it executes\[[71](https://arxiv.org/html/2608.06663#bib.bib145)\]\. But the recipe’s central risk is committing early to a plan the agent cannot actually carry out: Plan\-and\-Act starts from the observation that LLMs are not inherently trained to produce accurate plans, and trains a Planner separately from an Executor using synthetic data that annotates ground\-truth trajectories with the feasible plans that would have produced them — teaching plan generation directly, rather than hoping a single model’s general capability transfers to it — which recovers much of the gap on long\-horizon web\-navigation tasks\[[23](https://arxiv.org/html/2608.06663#bib.bib233)\]\. PIVOT goes further and treats the plan itself as an object to be refined against execution feedback rather than committed to once — plan, inspect, and evolve trajectories through repeated environment interaction — precisely because plans generated before any grounding routinely violate constraints or omit infeasible\-action checks that only surface once execution starts\[[126](https://arxiv.org/html/2608.06663#bib.bib1098)\]\. The same lesson appears from the harness\-design side: harnesses that decompose more aggressively are not uniformly better, because more elaborate decomposition can itself introduce misalignment between the guidance a step receives and what execution actually needs, so the right amount of decomposition is a property of the task’s uncertainty, not a free efficiency gain\[[91](https://arxiv.org/html/2608.06663#bib.bib1134)\]\.

Interleaving planning with execution\.The alternative — plan one step, observe, replan — trades a priori coherence for grounding\. PROMST shows that multi\-step tasks need feedback\- and heuristic\-informed prompt optimization that single\-step prompt tuning does not require, precisely because the impact of an individual step is hard to evaluate in isolation from what happens next\[[15](https://arxiv.org/html/2608.06663#bib.bib9)\]; ReAgent makes the interleaving explicit and reversible, adding backtracking to multi\-hop QA specifically because irreversible chain\-of\-thought accumulates errors across hops that only a revisable trajectory can undo\[[131](https://arxiv.org/html/2608.06663#bib.bib227)\]\. A sharp empirical question is exactly how much interleaving is warranted: directly testing full\-horizon planning \(commit to a complete plan, then execute\) against single\-step\-horizon planning \(interleave every action with fresh reasoning\) on data\-centric tool\-calling tasks, one study finds full\-horizon planning with on\-demand replanning matches step\-by\-step accuracy across the depths, breadths, and robustness levels tested — at 2\-3x fewer tokens — indicating the interleaved default’s assumption that eager step\-wise monitoring is necessary for adaptability does not hold for well\-defined data\-centric tasks, even though it is treated as the safe default\[[72](https://arxiv.org/html/2608.06663#bib.bib1084)\]\. This is a challenge to committing to interleaving by default, not a demonstration of the predictability\-conditioned split this section’s opening claims\. A cleaner test of that specific split is still open\. At the far end of this axis, when the action space itself becomes too large to plan over explicitly \(open\-ended real\-world environments\), one line of work argues the representation planning happens over must change — from planning with concrete actions to planning with higher\-level schemas that compress the combinatorially exploding action space back down to something a plan can range over\[[105](https://arxiv.org/html/2608.06663#bib.bib531)\]\.

Search\-based planning\.A third family keeps multiple candidate continuations alive and searches among them rather than committing to one\. The anchor for this family is Tree of Thoughts, which generalizes chain\-of\-thought from a single linear trace into a tree of partial solutions the model can evaluate and backtrack over\[[114](https://arxiv.org/html/2608.06663#bib.bib1555)\]; the within\-window variant remains the reference point against which later, tool\- and environment\-grounded search methods are described\. In our 2024\-2026 corpus this is the smallest slice \(search,n=9n=9\) but a conceptually distinct one: CheMatAgent integrates tree\-search over 137 external chemistry tools specifically because a single greedy tool\-call sequence is unreliable when the right tool depends on intermediate results not known in advance\[[97](https://arxiv.org/html/2608.06663#bib.bib342)\], and SELT combines self\-evaluation with tree search over decomposed sub\-tasks so that the search process itself, not just the final answer, is guided by a learned value signal\[[98](https://arxiv.org/html/2608.06663#bib.bib341)\]\. The progression is worth noting: Tree of Thoughts searches over*reasoning states*inside one context, while its successors search over*action sequences*whose outcomes only an environment can reveal — which moves the bottleneck from search strategy to the cost and fidelity of evaluating each branch\. Search\-based planning is the most compute\-hungry point on the spectrum and the literature increasingly asks whether the compute is well spent — a question §7 returns to when discussing benchmark protocol validity\.

World models for lookahead\.A fourth thread — closely related to search, since search needs something to search*over*— asks agents to maintain an explicit predictive model of the environment and plan by rolling it forward\. WALL\-E aligns a world model to real environment rules via rule learning specifically because an ungrounded world model’s predictions drift from what the environment actually does over a long rollout, and shows this alignment step recovers much of the value of model\-based planning that an ungrounded model loses\[[135](https://arxiv.org/html/2608.06663#bib.bib107)\]\. This drift is a recurring finding: a comparison of memory mechanisms in world models finds that the effective planning horizon of transformer\-based world models is bounded by their effective memory span, and that perceptual drift compounding across a long imagined rollout breaks the loop\-closure a plan depends on — connecting this thread directly to the memory literature of §4 and the compounding\-error theory of §8\[[46](https://arxiv.org/html/2608.06663#bib.bib639)\]\. Applied work shows both the promise and the constraint: MobileDreamer builds a generative sketch world model specifically to let a mobile GUI agent forecast action outcomes before committing to them, addressing agents that are otherwise purely reactive to the current screen and therefore blind on long\-horizon tasks\[[10](https://arxiv.org/html/2608.06663#bib.bib692)\], while ToolVerse scales agentic\-RL environments precisely because agents that reason well in compact, well\-defined scenarios do not automatically transfer that lookahead ability to large, diverse, dynamic ones\[[133](https://arxiv.org/html/2608.06663#bib.bib1391)\]\.

Planning that persists across tasks\.All four families above operate within a single attempt\. A smaller thread asks whether the planning capability itself can accumulate across episodes — the cross\-task\-persistent point on Axis 2 \(§2\.3\) as it applies to planning rather than to memory\. AgentEvolver treats the expensive parts of building an agent — constructing task datasets and exploring for useful trajectories — as things the system should bootstrap for itself rather than have hand\-supplied per task, on the argument that manually constructed datasets and undirected RL exploration are what make agent development costly and sample\-inefficient in the first place\[[120](https://arxiv.org/html/2608.06663#bib.bib604)\]\. SEAM takes a narrower and more architecturally explicit route: rather than retrieving past experience by similarity at inference time — which it argues introduces noise and latency — it trains a small executor\-specific adapter that*generates*a structured, instance\-tailored experience entry in one forward pass for a frozen executor, storing the accumulated experience in the adapter’s parameters and improving it after deployment from logged successful trajectories\[[53](https://arxiv.org/html/2608.06663#bib.bib763)\]\. The two mark the ends of a small design space — bootstrapping the whole training pipeline versus compiling reusable experience into a lightweight side module — and both make explicit an assumption the rest of this section leaves implicit: that a plan’s value can outlive the episode that produced it\.

Assessment\.Across the four within\-episode families, the same pattern recurs: methods that commit to a representation of the future early \(a full plan, an unrolled world\-model trajectory\) buy coherence at the price of fragility to whatever the plan did not anticipate, while methods that defer commitment \(interleaving, search\) buy robustness at the price of compute and sometimes at the price of the very global coherence long\-horizon tasks need\. No single point on the spectrum dominates; which point is appropriate is closest to being a measurable property of the environment’s uncertainty and the action space’s size, not a fixed methodological preference — which is why the field’s center of mass has moved from plan\-then\-execute defaults toward interleaved and search\-augmented methods as it has moved from short, well\-specified benchmarks toward open\-ended, long\-horizon ones \(§7\)\. The cross\-task thread is the least developed of the five and the one where the boundary with §4 is thinnest: once planning experience is stored and reused, the engineering problem starts to look like a memory problem, which is one reason the two categories interpenetrate as heavily as Figure 2 shows\. The plans this section describes are only as good as the information available when they are made and revised, which is the memory and context\-management problem taken up next\.

## 4\. Memory & Context Management

If planning decides what to do next, memory decides what information that decision is made with — and this is where the long\-horizon/long\-context/long\-term\-memory distinction of §2\.1 does the most work\. We organize this section by where information is carried:in\-context\(part of the live prompt\), anexternal store\(retrieved on demand\), orweights\(compiled into the policy via training\) — three points on a persistence\-fidelity trade\-off rather than three unrelated mechanisms \(contextn=103n=103,externaln=294n=294of thememorycategory; the size gap itself is a finding, discussed in the Assessment below — the weights tier is not separately tagged in the corpus, discussed qualitatively via thetrainingcategory’s overlap with this section\)\.

![Figure 4: The memory persistence-fidelity trade-off across three tiers -- context, external store, and weights -- with representative systems and relative fidelity/persistence for each.](https://arxiv.org/html/2608.06663v1/figures/figure4_memory_hierarchy.png)Figure 4: The memory persistence\-fidelity trade\-off across three tiers – context, external store, and weights – with representative systems and relative fidelity/persistence for each\.In\-context management\.The simplest way to preserve information is to keep it in the prompt, but a long\-horizon task eventually produces a trajectory longer than any context window, and even within budget, models do not use long contexts uniformly — information placed away from the ends of the context is used less reliably than information at the edges\[[61](https://arxiv.org/html/2608.06663#bib.bib1560)\]\. Agent\-specific work treats this as an active management problem rather than a passive capacity limit: Cognitive Workspace frames context management as active memory management with a finite functional budget, directly analogous to human working memory constraints, rather than as a fixed\-size buffer to be filled and discarded\[[6](https://arxiv.org/html/2608.06663#bib.bib428)\]; Git Context Controller manages an agent’s context the way a version control system manages a codebase — commits, branches, and merges of context state — because ad hoc truncation loses exactly the information a later step turns out to need\[[96](https://arxiv.org/html/2608.06663#bib.bib414)\]\. Several 2026 systems attack the trajectory\-length problem directly: Context\-Folding scales a long\-horizon agent by folding completed sub\-trajectories out of the live context once they are resolved\[[86](https://arxiv.org/html/2608.06663#bib.bib546)\], and ACON optimizes context compression specifically for the long\-horizon setting, showing that generic summarization compresses away details a later step needs, so the compression policy itself has to be aware of what downstream steps will require\[[39](https://arxiv.org/html/2608.06663#bib.bib517)\]\. The failure mode this thread converges on has recently been given a name — “context rot,” diagnosed and partially mitigated in long\-horizon search settings\[[103](https://arxiv.org/html/2608.06663#bib.bib1341)\]— and a blunt summary: context management is load\-bearing for whether an LLM agent’s plan survives contact with a long trajectory, not an implementation detail behind it\[[67](https://arxiv.org/html/2608.06663#bib.bib1306)\]\.

External memory\.The alternative to carrying everything in\-context is to write information to an external store and retrieve it when relevant\. This is by far the largest subcategory in the corpus \(external,n=294n=294vs\.context’s 103\), and the operating system metaphor that popularized the idea — MemGPT, treating an LLM’s limited context as working memory backed by a paged external store, the way an operating system backs limited RAM with disk\[[73](https://arxiv.org/html/2608.06663#bib.bib1558)\]— still organizes how most of this literature frames the problem\. A second anchor supplies the other half of the design vocabulary: Generative Agents paired a timestamped memory stream with a retrieval function scoring entries on recency, importance, and relevance, plus a reflection step that periodically synthesized higher\-level inferences from raw observations\[[75](https://arxiv.org/html/2608.06663#bib.bib1559)\]— establishing that an external store needs a*write*and*consolidation*policy, not only a read path, a point the interference findings below return to\. Recent systems differ mainly in what the external store looks like and how retrieval is scoped: A\-MEM structures agentic memory as a network of interlinked notes rather than a flat log, so that retrieval can follow associative links instead of similarity search alone\[[106](https://arxiv.org/html/2608.06663#bib.bib200)\]; graph\-based memory architectures generalize this further, arguing a single flat vector store under\-represents the relational structure a long\-running agent accumulates\[[109](https://arxiv.org/html/2608.06663#bib.bib797)\]\. Whether memory survives across sessions, not just within one long episode, is this subcategory’s clearest engineering marker of the cross\-task\-persistent point on Axis 2 \(§2\.3\): WebCoach gives web agents self\-evolving guidance carried across sessions specifically because within\-session learning is invisible to the next session’s fresh start\[[60](https://arxiv.org/html/2608.06663#bib.bib1536)\], and a scale\-conditioned evaluation protocol finds that fixed\-snapshot accuracy hides a real failure mode: reliability degrades as merely*irrelevant*sessions accumulate around the actually\-relevant evidence, well before the store approaches any hard capacity limit — the protocol reports the specific accumulation scale at which reliability crosses below target as a distinct diagnostic in its own right\[[79](https://arxiv.org/html/2608.06663#bib.bib1079)\]\. A closely related but distinct problem is interference from evidence that is not irrelevant but*outdated*: a benchmark evaluating memory under multi\-target interference in long\-horizon agent systems finds consistently low accuracy across seven representative systems \(vanilla long\-context LLMs, RAG, and memory\-augmented frameworks; average 27\.9%\), with questions that require aggregating multiple relevant facts singled out as especially hard even against that already\-low baseline, and degradation compounding further as facts are revised or updated by later context — interference from*change over time*, not merely from how much is stored, is the bottleneck\[[47](https://arxiv.org/html/2608.06663#bib.bib1152)\], and forgetting is correspondingly being explored as a first\-class design target rather than a failure to prevent — biologically\-inspired forgetting mechanisms compress or discard low\-value memories specifically to keep a growing store’s retrieval useful\[[95](https://arxiv.org/html/2608.06663#bib.bib742)\]\.

Compression and distillation into weights\.The third tier folds information into the model itself rather than an external artifact — the most persistent but least flexible point on the trade\-off, since updating it requires training rather than a write to a store\. MemPO frames this as self\-memory policy optimization for long\-horizon agents, training the policy itself via reinforcement learning to autonomously decide what to retain or summarize based on downstream usefulness — the skill of curation becomes a trained capability baked into the weights, rather than an external module the policy merely calls\[[52](https://arxiv.org/html/2608.06663#bib.bib869)\]; co\-evolving decision and skill\-bank agents blend the tiers explicitly, pairing a fast\-changing external skill bank with a more slowly updated decision policy so that frequently reused skills eventually migrate from retrieval into the policy itself\[[100](https://arxiv.org/html/2608.06663#bib.bib1022)\]\. This tier is the most durable against context\-window limits by construction, and correspondingly the least auditable — a concern §8 returns to\.

A cross\-cutting risk: memory as an attack surface\.Because external and weight\-level memory both persist information an attacker does not control at write time, several 2026 papers treat agent memory itself as a security boundary rather than only a capability: conversational interaction can plant a stealthy trojan directly into an agent’s stored memory\[[93](https://arxiv.org/html/2608.06663#bib.bib1203)\], and a forensic trajectory signature for detecting memory\-poisoning after the fact reaches high discrimination in a preregistered evaluation, but a follow\-up finds the same signature also fires on benign memory\-grounded behavior — a reminder that a detector’s headline accuracy and its practical false\-positive rate under realistic use are separate questions this literature is only beginning to ask\[[48](https://arxiv.org/html/2608.06663#bib.bib1338)\]\. This connects the memory literature directly to the oversight problem of §8: a memory store is exactly the kind of persistent, agent\-controlled state that is hard for a human overseer to audit continuously\.

Assessment\.Theexternal\-to\-contextsize ratio in the corpus \(294 vs\. 103\) is itself informative: as of this writing, the field has moved decisively toward treating memory as a storage\-and\-retrieval engineering problem rather than a context\-budgeting one, likely because external stores scale independently of a fixed context window while in\-context management schemes are ultimately still bounded by it\. But the persistence\-fidelity trade\-off runs through all three tiers regardless of implementation: context is faithful but bounded, external stores are unbounded but suffer retrieval and interference costs that grow with scale, and weights are durable but require training to update and are the hardest tier to audit or correct after the fact\. None of the three eliminates the trade\-off; they only choose where on it to sit\. Whichever tier a system chooses, the resulting information still has to be acted on inside a runtime loop that can fail and needs to recover — the subject of §5\.

## 5\. Execution Control & Recovery

Planning and memory supply what to do and what to do it with; execution control is the runtime loop that actually does it, and — the recurring claim of this section — how well a long task goes depends more on that loop’s ability to notice and recover from failure than on the per\-step quality of the underlying model\. This is the corpus’s largest category \(execution,n=584n=584\), organized into three subcategories that sit at increasing scale: the single\-agent act–observeloopthat turns a model into an agent at all,orchestrationacross multiple agents or sub\-agents \(n=338n=338, the dominant mode by volume\), andrecoveryvia self\-correction and replanning when the loop’s output is wrong \(n=245n=245\)\.

The loop itself\.The interleaved reason\-act\-observe cycle — think, take an action, read back the environment’s response, repeat — is now so assumed a substrate for LLM agents that the corpus contains almost no papers that treat the loop itself as a subject rather than infrastructure \(loop,n=1n=1\); the pattern’s near\-total absorption into “how agents just work” is itself informative about how settled this layer has become since ReAct first proposed it as an explicit alternative to acting without intermediate reasoning traces\[[115](https://arxiv.org/html/2608.06663#bib.bib1552)\]\. What the corpus does discuss is how to make the loop*safe*to run for a long time unsupervised: a practitioner\-facing guide to architecting resilient plan\-then\-execute agents treats the loop’s exception paths — what happens when a tool call fails, times out, or returns something the plan did not anticipate — as first\-class design surface rather than an afterthought bolted onto a working happy path\[[19](https://arxiv.org/html/2608.06663#bib.bib463)\]\.

Orchestration across agents\.The dominant way the field has scaled beyond a single loop is horizontally: split a task across multiple agents or sub\-agents rather than deepen one agent’s own loop\. Two anchors bracket what “delegation” came to mean: Toolformer had a model learn, in a self\-supervised way, when to call an external API mid\-generation — delegation to a*tool*\[[78](https://arxiv.org/html/2608.06663#bib.bib1556)\]— while HuggingGPT used a language model as a controller that plans a task and dispatches subtasks to other specialist models, delegation to a*model*\[[80](https://arxiv.org/html/2608.06663#bib.bib1557)\]\. Present\-day orchestration generalizes the latter: the delegate is another agent with its own loop\. This ranges from domain\-specialized multi\-agent frameworks — for requirements engineering\[[35](https://arxiv.org/html/2608.06663#bib.bib399)\], sentiment analysis\[[107](https://arxiv.org/html/2608.06663#bib.bib441)\], penetration testing\[[63](https://arxiv.org/html/2608.06663#bib.bib472),[58](https://arxiv.org/html/2608.06663#bib.bib647)\], optical\-network operations\[[127](https://arxiv.org/html/2608.06663#bib.bib530)\], and incident response with an explicit deterministic\-decision\-support goal\[[22](https://arxiv.org/html/2608.06663#bib.bib613)\]— to general\-purpose compositional frameworks for a single embodiment, such as Agent S2’s generalist\-specialist split for computer\-use agents\[[3](https://arxiv.org/html/2608.06663#bib.bib251)\]and OmegaUse’s general\-purpose GUI agent\[[124](https://arxiv.org/html/2608.06663#bib.bib756)\]\. A second thread treats orchestration itself, not any one domain, as the object of study: the first large\-scale empirical study of testing practices across 39 open\-source agent frameworks and 439 agentic applications finds testing effort concentrated almost entirely on the deterministic scaffolding around a model — tools and workflows absorb over 70% of testing effort — while the model\-driven planning component itself receives under 5% and prompts under 1%, an inversion that leaves exactly the least deterministic, most agentic part of the system least tested\[[28](https://arxiv.org/html/2608.06663#bib.bib490)\]; a two\-dimensional framework separates an agent’s cognitive function from its execution topology specifically because neither axis alone disambiguates architecturally distinct systems — the same orchestration topology can implement patterns with very different failure modes\[[30](https://arxiv.org/html/2608.06663#bib.bib913)\]\. The governance literature has started treating agentic deployment as a policy object in its own right, from two different angles: “Governing AI Agents” applies the economic theory of principal\-agent problems to argue that conventional governance tools — monitoring, incentive design, enforcement — may not transfer to AI agents that act at a speed and opacity ordinary agency relationships did not anticipate\[[42](https://arxiv.org/html/2608.06663#bib.bib165)\], while the AI Agent Index, a public database of deployed agentic systems, finds developers document capabilities and applications far more thoroughly than they document safety and risk\-management practices\[[12](https://arxiv.org/html/2608.06663#bib.bib186)\]— a disclosure gap this section shares with the oversight discussion in §8\.

Recovery: self\-correction and replanning\.The third subcategory is the loop’s error\-handling layer — noticing that an action or a reasoning step was wrong and repairing it before it compounds\. The pattern’s anchor is Reflexion, which converts a failed attempt into natural\-language self\-feedback stored in an episodic buffer and conditions the next attempt on it — verbal reinforcement in place of a gradient update\[[83](https://arxiv.org/html/2608.06663#bib.bib1553)\]\. Early optimism that models could reliably self\-correct their own reasoning without external feedback was tempered by a widely cited finding that intrinsic self\-correction — correction using only the model’s own signal, with no external verifier or ground truth — does not reliably improve reasoning accuracy and can degrade it\[[31](https://arxiv.org/html/2608.06663#bib.bib1573)\]; the corpus’s own trajectory traces this tension rather than resolving it\. Several 2024\-2025 papers argue for a more qualified intrinsic capability under narrower conditions — self\-correction of single\-utterance perturbed reasoning\[[84](https://arxiv.org/html/2608.06663#bib.bib364)\], correction with explicit key\-condition verification\[[101](https://arxiv.org/html/2608.06663#bib.bib31)\]— while a 2025 decomposition of self\-correction into detection, localization, and correction sub\-capabilities surfaces a genuinely counter\-intuitive finding it calls the “accuracy\-correction paradox”: the weaker model in a three\-model comparison \(GPT\-3\.5, 66% base accuracy\) corrects its own errors intrinsically at a*higher*rate \(26\.8%\) than the strongest model \(94% base accuracy, 16\.7%\), and error\-detection rate does not predict correction success either\. The proposed explanation — an “Error Depth Hypothesis” — is that stronger models make fewer but structurally deeper errors that intrinsic self\-correction is specifically bad at reaching, which would mean the anchor finding above is not a capability gap current models simply haven’t crossed yet, but a pattern that could get*worse*, not better, as models improve\[[54](https://arxiv.org/html/2608.06663#bib.bib669)\]\. Where self\-correction is externally grounded, the picture is more positive: CSC\-SQL uses corrective self\-consistency, checking multiple candidates against each other rather than relying on one model’s unaided judgment, to improve text\-to\-SQL reliability\[[81](https://arxiv.org/html/2608.06663#bib.bib298)\], and SHIELDA gives agentic workflows structured exception handling analogous to a programming language’s exception model, so that failures are caught and routed rather than silently propagated\[[132](https://arxiv.org/html/2608.06663#bib.bib431)\]\. At the level of a full trajectory rather than a single step, task\-decoupled planning replaces one monolithic reasoning history spanning every sub\-task with a directed acyclic graph of scoped sub\-goals, each with its own confined context, precisely because a shared history lets an error made on one sub\-task propagate into otherwise\-unrelated decisions and makes recovery correspondingly expensive to localize\[[55](https://arxiv.org/html/2608.06663#bib.bib706)\]\.

Recovery that accumulates\.A distinct move within the recovery literature is to treat each failure\-and\-repair episode as a durable asset rather than a one\-off correction — the cross\-task\-persistent locus \(§2\.3\) applied to execution\. The anchor here is Voyager, which accumulated a library of executable skills in an open\-ended Minecraft setting so that capabilities acquired solving one task became available for the next\[[92](https://arxiv.org/html/2608.06663#bib.bib1554)\]\. ViReSkill pairs vision\-grounded replanning with a skill memory: when execution fails, the replanner generates a new action sequence conditioned on the current scene, and when it succeeds, the resulting plan is retained for reuse — so the system’s competence on a task family grows with the number of failures it has worked through\[[37](https://arxiv.org/html/2608.06663#bib.bib512)\]\. SEAgent makes the same bet on a larger scale for computer\-use agents, letting an agent autonomously master unfamiliar software through iterative trial\-and\-error on auto\-generated tasks, precisely because human\-labeled demonstrations do not exist for novel or specialized applications\[[87](https://arxiv.org/html/2608.06663#bib.bib424)\]\. Both depend on something §5’s within\-episode recovery methods do not need: a reliable judgment about which past attempts were actually successes, since a skill memory that accumulates plausible\-looking failures degrades rather than improves — which routes the problem back to the evaluation question of §7\.

Assessment\.The subcategory sizes are themselves a claim: orchestration outnumbers recovery roughly 1\.4 to 1, and both dwarf the single\-loop subcategory, suggesting the field has invested far more engineering effort in*scaling out*\(more agents, more structure around them\) than in*hardening*any one agent’s own error\-correction — even though the self\-correction evidence above suggests hardening is where the harder unsolved problem sits\. This is consistent with this section’s opening claim: what separates a system that completes a long task from one that does not is decreasingly a property of the underlying model and increasingly a property of the harness wrapped around it — how failures are caught, how many agents share the load, and how aggressively the system attempts to fix its own mistakes before they compound\. The cross\-task\-persistent accumulation thread above sharpens this same attribution problem rather than sidestepping it: a skill memory only compounds capability if the system’s judgment about which past attempts succeeded is itself trustworthy, which routes straight back to the evaluation question of §7 and means accumulation is not a way around the harness\-versus\-model question so much as a second place it resurfaces\. Whether the harness handling all of this was itself trained to handle long horizons, or is purely inference\-time scaffolding bolted onto a model trained for something else, is the question §6 takes up\.

## 6\. Training for Long Horizons

The previous two sections describe scaffolding wrapped around a trained model\. This section asks what happens when the model itself is trained with long horizons in mind, and its organizing observation is simple: as horizon grows, outcome\-only supervision — one reward at the very end of a long trajectory — grows sparser and noisier per step, which is exactly why process\-level supervision and credit assignment have become the field’s central technical problem rather than a footnote to reinforcement learning as usual \(rln=130n=130,supervisionn=37n=37of thetrainingcategory\)\.

Credit assignment under sparse, delayed reward\.When a single trajectory reward must be distributed back across dozens of intermediate actions, naive uniform credit assignment systematically mis\-attributes success and failure to steps that did not cause them\. Segment Policy Optimization addresses this directly by assigning credit at the level of trajectory segments rather than whole trajectories or individual tokens, on the premise that neither granularity extreme matches where causal responsibility actually lives\[[26](https://arxiv.org/html/2608.06663#bib.bib319)\]; a graph\-based approach goes further, arguing that attribution restricted to a single trajectory’s own linear order — even at segment granularity — misses credit relationships that only appear across a graph of related trajectories\[[17](https://arxiv.org/html/2608.06663#bib.bib1183)\]\. Selective eligibility traces revisit a much older RL idea specifically to avoid the opposite failure of uniform credit assignment: spreading credit equally over every step of a long trajectory dilutes the signal precisely where it is most informative\[[69](https://arxiv.org/html/2608.06663#bib.bib1074)\]\. Some of this thread ports credit\-assignment machinery into domains distant from its origin — multi\-granularity intertemporal credit assignment for long\-horizon emotional\-support dialogue, combining immediate and delayed credit from a shared potential function over dialogue state\[[125](https://arxiv.org/html/2608.06663#bib.bib887)\], fair credit assignment for memory\-augmented agents where the credit a memory\-write step deserves is entangled with whichever later step actually uses that memory\[[108](https://arxiv.org/html/2608.06663#bib.bib1166)\]— evidence that credit assignment has become a general\-purpose long\-horizon primitive rather than a technique specific to any one task family\. A comprehensive empirical recipe for reinforcement learning on long\-horizon tool\-using agents decomposes the agentic\-RL design space along five axes — reward shaping, model scale, data composition, algorithm choice, and environmental stability — and is explicit that which choice is best is itself scale\-dependent \(smaller models benefit from staged rewards and extra exploration; larger models converge faster with simpler dense rewards\), meaning there is no single best recipe outcome\-only RL alone would reveal\[[99](https://arxiv.org/html/2608.06663#bib.bib926)\]\.

Process reward models and generative supervision\.Where credit assignment asks*how to attribute*a sparse outcome signal, process reward models ask whether a*denser*signal can be trained directly from step\-level supervision instead of derived post hoc\. The reference result for this direction is Lightman et al\.’s comparison of process\- against outcome\-supervised reward models on the MATH dataset, which reported that supervising each reasoning step significantly outperformed supervising only the final answer — but required 800,000 step\-level human feedback labels \(released as PRM800K\) to train the better reward model\[[57](https://arxiv.org/html/2608.06663#bib.bib1572)\]\. That pairing is the inheritance this subsection works against: dense supervision demonstrably helps, at an annotation cost no agentic setting can pay per task\. Nearly everything below is an attempt to obtain step\-level density without step\-level human labels\. Entropy\-regularized process reward modeling directly targets a distinct failure mode of naively trained PRMs — overconfident, poorly calibrated step scores — by regularizing against it during training\[[123](https://arxiv.org/html/2608.06663#bib.bib142)\], and GroundedPRM grounds step\-level rewards in explicit tree\-guided search plus a fidelity check, precisely because ungrounded PRM training tends to reward plausible\-looking steps rather than steps that are actually correct\[[128](https://arxiv.org/html/2608.06663#bib.bib556)\]\. One striking observation ties this subsection back to §6’s opening claim about outcome supervision: GRPO — an algorithm introduced and typically described purely in outcome\-reward terms — is shown to implicitly perform process\-reward\-like credit attribution as a side effect of its group\-relative normalization, suggesting the boundary between “outcome” and “process” supervision is less an architectural choice than a question of where in a training pipeline step\-level credit ends up being computed\[[85](https://arxiv.org/html/2608.06663#bib.bib494)\]\. Reward density is also being pushed toward the long\-horizon, agent\-specific setting rather than borrowed unmodified from short\-horizon math/code reasoning: Think\-RM extends long\-horizon reasoning into the generative\-reward\-model itself, so that the judge doing the rewarding is trained with the same long\-horizon considerations as the policy it supervises\[[29](https://arxiv.org/html/2608.06663#bib.bib303)\], AgentPRM builds step\-wise promise\-and\-progress signals specific to agent trajectories rather than reasoning chains\[[102](https://arxiv.org/html/2608.06663#bib.bib599)\], and SWE\-TRACE combines rubric\-based process rewards with heuristics specifically to make long\-horizon software\-engineering\-agent training tractable\[[27](https://arxiv.org/html/2608.06663#bib.bib1003)\]\. A separate cautionary finding shows implicit process signal is not always benign even when it exists: one paper documents agentic RL for search actively misaligning a model’s instruction\-following behavior as an unintended side effect of the reward it optimizes\[[112](https://arxiv.org/html/2608.06663#bib.bib566)\]— a different mechanism from GRPO’s implicit\-PRM structure above, but a reminder that reward shaping in agentic RL has effects beyond the task metric it targets, a finding this section flags rather than resolves, since it depends heavily on the specific reward and environment involved\.

Boundary with the self\-improvement literature\.Much of the machinery in this section — reward models, RL fine\-tuning on model\-generated trajectories, iterative policy improvement — overlaps mechanically with the recursive\-self\-improvement literature surveyed in a companion paper\[[13](https://arxiv.org/html/2608.06663#bib.bib1574)\]\. The line we draw is one of purpose, not mechanism: that survey treats these loops as an end in themselves — a system that trains on its own outputs, evaluates its own quality signal, or improves its own research process, with the interesting question being*whether the loop closes*\. Here, the same techniques are training a policy to act well over more steps, with a fixed external task\-completion criterion the training never gets to redefine — the interesting question is*how far one trained policy can be pushed on longer tasks*, not whether the training loop itself is self\-referential\. A single\-agent RL run that uses self\-generated rollouts to train a longer\-horizon tool\-using policy sits, by this criterion, inside this section; a system that additionally modifies its own reward model, training procedure, or harness as part of the same loop crosses into the companion survey’s territory\. Papers that live on this boundary — e\.g\., Q\-Evolve, which unifies automatic process\-reward labeling with policy learning in a self\-evolving, in\-distribution RL loop\[[129](https://arxiv.org/html/2608.06663#bib.bib1544)\]— are treated here strictly as instances of longer\-horizon policy training, leaving the self\-referential\-loop question to the companion paper\.

Assessment\.Therl\-to\-supervisionratio \(130 to 37\) suggests the field has invested more effort in*assigning*credit within a fixed reward signal than in*redesigning*the reward signal’s granularity outright — the cheaper intervention, and consistent with credit assignment’s portability across the very different domains cited above\. But the GRPO finding above complicates a clean read of that ratio: if outcome\-level algorithms are already doing implicit process\-level work, the two subcategories are less separate research programs than two vantage points on the same underlying problem — training a policy that gets useful signal from every step of a long trajectory, not only its end\. Whether that training actually produces agents that are more reliable over long horizons, and how we would know, is the measurement problem taken up next\.

## 7\. Evaluation & Measurement

Every claim in §§3–6 is only as credible as the benchmark it was measured against, and this section’s organizing observation is that agentic benchmarks have moved from single\-step QA toward hours\-long, real\-environment tasks faster than the methodology for validating them has matured — so that a large and increasingly self\-aware slice of this literature is now about whether the benchmarks measure what they claim to \(evaluation, entirely taggedbenchmarkin our taxonomy,n=114n=114, plus the SWE\-bench/GUI benchmark ecosystem that recurs throughoutexecution\)\. All benchmark numbers below are dated at first mention, since public leaderboards shift on a timescale of months\.

A case study in benchmark evolution: SWE\-bench\.SWE\-bench\[[34](https://arxiv.org/html/2608.06663#bib.bib1565)\]— resolve real GitHub issues given the full repository as context — has become the closest thing long\-horizon agent evaluation has to a common currency, and its evolution is a microcosm of this section’s argument\. The original benchmark and its immediate successors extended*coverage*: SWE\-bench\-java\[[119](https://arxiv.org/html/2608.06663#bib.bib86)\]and Multi\-SWE\-bench\[[118](https://arxiv.org/html/2608.06663#bib.bib254)\]add languages beyond Python, SWE\-bench Multimodal asks whether solutions generalize to visual software domains the text\-only original cannot capture\[[110](https://arxiv.org/html/2608.06663#bib.bib100)\], and SWE\-Bench Pro raises the difficulty to enterprise\-scale, long\-horizon problems explicitly because the original’s scope undersells the length of task real software engineering requires\[[20](https://arxiv.org/html/2608.06663#bib.bib483)\]\. A second, larger wave interrogates*validity*rather than coverage\. SWE\-bench\+ manually inspects successful patches and finds 32\.67% involve solution leakage \(the fix was already present in the issue report or comments\) and a further 31\.08% pass only because the test suite is too weak to verify correctness — filtering both out drops one leaderboard\-topping system’s resolution rate from 12\.47% to 3\.97%\[[5](https://arxiv.org/html/2608.06663#bib.bib105)\]; “Are ‘Solved Issues’ in SWE\-bench Really Solved Correctly?” corroborates the second half of that finding independently, showing passing patches diverge behaviorally from the human\-written ground truth in nearly 30% of cases even when tests pass\[[94](https://arxiv.org/html/2608.06663#bib.bib240)\]; SWE\-MERA responds by building a dynamic, continuously\-updated benchmark instead of a frozen one, arguing that the contamination and weak\-test problems SWE\-bench\+ documents are inherent to any benchmark that stops collecting new issues after release\[[1](https://arxiv.org/html/2608.06663#bib.bib397)\]; and “The SWE\-Bench Illusion” shows state\-of\-the\-art LLMs sometimes succeed by recalling memorized repository content rather than reasoning about the issue in front of them\[[56](https://arxiv.org/html/2608.06663#bib.bib354)\]— a concern sharpened further by a study asking directly whether SWE\-Bench Verified tests agent ability or model memory, and finding evidence for the latter in frequently\-benchmarked models\[[76](https://arxiv.org/html/2608.06663#bib.bib650)\]\. A third strand asks whether the leaderboards built on top of the benchmark are themselves trustworthy: dissecting SWE\-Bench’s public leaderboards finds systematic differences in how submitters profile and report results that complicate cross\-system comparison\[[65](https://arxiv.org/html/2608.06663#bib.bib367)\], corroborated by a follow\-up comprehensive study of the same two leaderboards\[[66](https://arxiv.org/html/2608.06663#bib.bib793)\]\. Most recently, a trajectory\-level diagnostic \(TRAJEVAL\) shows that even where capable models localize the right code, they still fail after reaching it — “coherence collapse” — meaning Pass@1 alone actively misdiagnoses*why*the remaining 30–35% of issues go unsolved\[[40](https://arxiv.org/html/2608.06663#bib.bib933)\]\. Best\-practice guidance distilled from across this benchmark family generalizes the lesson beyond SWE\-bench: across a wider set of agentic benchmarks, task\-setup and reward\-design flaws \(SWE\-bench Verified’s insufficient test cases are exhibit one; a separate example findsτ\\tau\-bench\[[113](https://arxiv.org/html/2608.06663#bib.bib1569)\]— a benchmark for tool\-agent\-user interaction under domain policies, notable for scoring reliability across repeated trials of the same task rather than single\-shot success — counts empty responses as successes\) can over\- or under\-estimate reported performance by up to 100% in relative terms, and a resulting checklist reduced measured overestimation on one complex benchmark by 33% when applied\[[136](https://arxiv.org/html/2608.06663#bib.bib382)\]\.

GUI and computer\-use benchmarks\.A parallel ecosystem evaluates agents that operate graphical interfaces rather than repositories\. WebArena\[[134](https://arxiv.org/html/2608.06663#bib.bib1566)\]and OSWorld\[[104](https://arxiv.org/html/2608.06663#bib.bib1570)\]established the realistic\-environment paradigm for web and full\-desktop tasks respectively; AgentBench\[[62](https://arxiv.org/html/2608.06663#bib.bib1567)\]and GAIA\[[68](https://arxiv.org/html/2608.06663#bib.bib1568)\]cover a wider span of agentic environments and general\-assistant tasks\. The corpus’s own supplement here is dense with reliability findings that echo the SWE\-bench pattern in a different embodiment: OS\-Harm evaluates the safety, not just the task success, of computer\-use agents specifically because prior work “largely overlooked” harm potential while chasing capability\[[44](https://arxiv.org/html/2608.06663#bib.bib360)\]; AndroidControl\-Curated finds that even top\-scoring GUI agents plateau around 60% on AndroidControl not because the task is unsolvable but because benchmark noise — ambiguities and factual errors in the benchmark itself — caps the achievable score below what the underlying capability would otherwise allow\[[49](https://arxiv.org/html/2608.06663#bib.bib568)\]; and a direct study “On the Reliability of Computer Use Agents” asks the sharpest version of this question: if an agent succeeds at a task once, what stops it from succeeding again on an unchanged retry, and attributes the answer to three factors: stochasticity during execution, ambiguity in how the task itself is specified \(room for divergent\-but\-valid interpretations\), and run\-to\-run variability in the agent’s own behavior — reliability, in other words, is not purely a property of the model, but of execution, task specification, and agent behavior jointly\[[25](https://arxiv.org/html/2608.06663#bib.bib1016)\]\. Bilingual and cross\-platform coverage is a growing concern rather than an afterthought: macOSWorld extends the realistic\-environment paradigm to a major OS the original benchmarks omit\[[111](https://arxiv.org/html/2608.06663#bib.bib331)\], and WindowsWorld argues that single\-application benchmarks like OSWorld understate the difficulty of realistic, cross\-application professional workflows\[[51](https://arxiv.org/html/2608.06663#bib.bib1044)\]\.

Measuring horizon directly: time\-horizon metrics\.Rather than asking whether an agent solves a fixed task, a distinct line of work asks how*long*a task an agent can complete at a given reliability level, measured in human\-time\-to\-complete — an attempt to put a single number on the “how long a horizon can this system carry” question this whole survey is organized around\. The most visible instance of this measurement approach reports that the task length frontier models can complete at a fixed success rate has grown at a measured, roughly exponential rate over recent model generations\[[45](https://arxiv.org/html/2608.06663#bib.bib1571)\]\. We flag this explicitly as*measured and reported*, not as a law: doubling\-type extrapolations of this kind are exactly the sort of overclaim this survey’s language calibration \(§2\.2\) commits to avoiding, and the metric’s own assumptions — what counts as “the same” task across models of different eras, and whether human\-time\-to\-complete on a benchmark task translates to human\-time\-to\-complete on the messier tasks §1 opens with — are contested rather than settled\. We treat this metric, and the critical literature interrogating agentic\-benchmark protocol validity more broadly\[[136](https://arxiv.org/html/2608.06663#bib.bib382)\], as a first\-class thread rather than a premise: a benchmark from the SWE\-bench and GUI ecosystems above showing a capability plateau is only informative to the extent the benchmark’s own construction is trustworthy, which is precisely what §7’s second and third paragraphs show cannot be assumed by default\.

![Figure 5: Human time-to-complete one task instance across the small set of benchmarks that report it directly, log scale. Dots are point estimates; bars are reported or illustrative ranges (see caption sources). Sourced entirely from primary papers -- artifacts/benchmark_durations.csv.](https://arxiv.org/html/2608.06663v1/figures/figure5_benchmark_duration.png)Figure 5: Human time\-to\-complete one task instance across the small set of benchmarks that report it directly, log scale\. Dots are point estimates; bars are reported or illustrative ranges \(see caption sources\)\. Sourced entirely from primary papers – artifacts/benchmark\_durations\.csv\.Figure 5 makes a point about the field’s measurement maturity by its own sparseness: of the roughly fifteen well\-known agentic benchmarks we checked against their own papers, only GAIA reports a clean, level\-by\-level human\-time\-to\-complete statistic; METR’s own composite task suite states the band over which model success degrades rather than the suite’s literal range; and SWE\-Bench Pro states only a qualitative “hours to days\.” SWE\-bench itself, WebArena, and OSWorld — three of the most widely used agentic benchmarks in this literature — report no human\-time baseline at all in their own papers\. A field that wants METR\-style time\-horizon measurement to generalize across benchmarks needs more benchmarks to report the human\-time statistic that measurement depends on, not just more benchmarks\.

Evaluating what persists across tasks\.Every benchmark discussed so far scores an agent on independent task instances, which by construction cannot see whether an agent got better at the*next*task for having done the previous one\. A small but distinct thread builds benchmarks for exactly that\. SWE\-Bench\-CL reorganizes SWE\-bench Verified’s issues into chronologically ordered per\-repository sequences, so that experience accumulation, knowledge transfer across tasks, and resistance to catastrophic forgetting become directly measurable rather than incidental\[[36](https://arxiv.org/html/2608.06663#bib.bib353)\]\. The ELL framework pairs a lifelong\-learning agent architecture with a benchmark built on the same premise — that the interesting question for open\-ended agents is continuous growth through interaction, not performance on a static task set\[[9](https://arxiv.org/html/2608.06663#bib.bib443)\]\. EvoAgentBench sharpens the target further, arguing that what should transfer is*procedural*— reusable searching, debugging, and verification routines extracted from execution traces — and that neither single\-episode agent benchmarks nor information\-retention memory benchmarks isolate it\[[24](https://arxiv.org/html/2608.06663#bib.bib1460)\]\. This thread is small \(fewer than a dozen papers in our corpus\), which is itself notable: the systems literature of §§3\-5 is full of agents that claim to improve across sessions, but the evaluation literature has only recently begun building instruments that could confirm or refute those claims independently\.

Assessment\.The evaluation literature’s shape mirrors the training literature’s from §6: just as outcome\-only reward grows uninformative as horizon grows, outcome\-only benchmark scores \(did the patch pass the tests? did the agent reach the goal screen?\) grow less informative as the tasks being measured get longer and the ways to superficially satisfy a check without solving the underlying problem multiply\. The field’s response — data purification, leakage audits, trajectory\-level diagnostics, reliability\-under\-retry studies, direct human\-time\-based horizon measurement — is evaluation’s version of the move from outcome to process supervision in §6, and the two are mutually entangled in a way this section can describe but not resolve: better benchmarks are needed to tell whether long\-horizon training methods work, and long\-horizon training methods are increasingly what benchmark designers use to probe where a benchmark’s own construction breaks\. Whether that entanglement \(§9 examines what it does and does not license concluding\), and the degradation patterns both this section and §6 keep surfacing, admit any general theory is the question §8 takes up\.

## 8\. Foundations, Limits & Safety

The preceding sections describe engineering responses to a common pressure: as horizon grows, small per\-step error rates compound, and unattended systems accumulate risk that no single step reveals\. This section asks how far the field has gotten toward a general account of that compounding, and toward the oversight problem it creates for agents that run for a long time with no human watching every step\. It is this survey’s smallest technical category by design \(n=103n=103, after a targeted supplement specifically because the seed harvest badly under\-covered it, §2\.2\), and its content is disclosed as more theoretical, diagnostic, and safety\-oriented than empirically settled\.

A simple model of compounding error, and where it breaks\.The textbook picture of long\-horizon degradation is straightforward: if per\-step error isϵ\\epsilonand errors compound independently, success probability falls off exponentially in the number of steps — a model implicit in the “why should a 100\-step task be so much harder than a 10\-step task” intuition motivating this whole survey\. What the corpus’s diagnostic literature shows repeatedly is that real degradation curves deviate from this simple picture in informative ways, on both sides\. Some deviations are worse than independent compounding predicts: “Strained Coherence” identifies a measurable pre\-failure signal in coding\-agent execution trajectories — trajectories that are about to fail show detectable strain before the failure itself manifests, meaning the error process is not memoryless the way the simple model assumes\[[74](https://arxiv.org/html/2608.06663#bib.bib1504)\], and “Coherence Collapse” \(§7\) shows failures cluster at a specific stage of a trajectory rather than distributing uniformly across steps\[[40](https://arxiv.org/html/2608.06663#bib.bib933)\]\. Other deviations complicate the picture differently: Vending\-Bench evaluates long\-term coherence directly \(can an autonomous agent run a simulated vending\-machine business over an extended horizon — runs exceeding 20M tokens — without losing track of its own state\) and finds high variance rather than smooth decay: capable models turn a profit in most runs, but every model has some runs that derail into a “meltdown” loop from which they rarely recover; notably, derailment shows no clear correlation with the context window filling up, suggesting these breakdowns are not simply a memory\-capacity story\[[8](https://arxiv.org/html/2608.06663#bib.bib1446)\]\. Together these findings support a modest but real conclusion: independent per\-step error compounding is a useful null model, not an empirically confirmed law, but the alternative it points toward is not gentler degradation — it is bimodal outcomes \(fine, or catastrophically derailed\) whose trigger the field cannot yet reliably predict\.

Goal drift and misalignment over long horizons\.A related but distinct failure mode is not that an agent’s actions become less correct, but that its objective silently shifts\. A technical report evaluating goal drift in language model agents documents this directly as a distinct phenomenon from accuracy decay\[[7](https://arxiv.org/html/2608.06663#bib.bib1489)\], and “Governance Decay” shows a specific, insidious mechanism for it: context compaction — the same compression §4 already treats as an unresolved persistence\-fidelity trade\-off, not a solved one, even before its safety implications are considered — can silently erase the safety constraints an agent was given at the start of a long trajectory, precisely because a compression policy optimized for task\-relevant information has no reason to preserve constraints that never come up again until they are violated\[[14](https://arxiv.org/html/2608.06663#bib.bib1470)\]\. Multi\-agent settings compound this further: “The Coming Crisis of Multi\-Agent Misalignment” argues that AI alignment work has not kept pace with the dynamic, social nature of misalignment that emerges specifically from*interaction*between agents rather than from any single agent’s own objective\[[11](https://arxiv.org/html/2608.06663#bib.bib1496)\]\. A formal treatment gives that interaction\-driven concern a specific mechanism: under a Bayesian model of automated multi\-agent workflows with weak per\-workflow evidence, each agent’s generic pretraining prior over its own behavior dominates the comparatively weak task\-specific signal in its workflow prompt — a failure the paper names “posterior collapse” by analogy to the same term in variational inference — with the practical consequence that distinct agents assigned distinct roles converge on nearly identical actions, disregarding the role distinctions the workflow was designed around; the paper argues this can be corrected only by injecting context\-specific evidence back into each agent’s belief update, not by better utility design alone\[[117](https://arxiv.org/html/2608.06663#bib.bib1493)\]\. Deception is a further specific concern this literature has started to measure directly rather than only theorize about: “Cheap Talk, Empty Promise” documents frontier LLMs breaking public commitments for self\-interest in controlled settings\[[82](https://arxiv.org/html/2608.06663#bib.bib1486)\], and a werewolf\-game study operationalizes deception and falsehood detection in a controlled multi\-agent environment specifically because naturalistic settings make ground truth about an agent’s beliefs hard to establish\[[2](https://arxiv.org/html/2608.06663#bib.bib1436)\]\.

Oversight for long\-running, unattended agents\.If an agent runs unattended for a long horizon, the standard human\-in\-the\-loop safety pattern — review every action before it executes — is exactly what “long\-running” and “unattended” preclude, which is why this sub\-thread treats oversight design as a first\-order engineering problem rather than a policy afterthought\. Tiered Agentic Oversight proposes a hierarchical multi\-agent oversight structure modeled on clinical hierarchy \(nurse, physician, specialist\) for safety\-critical healthcare settings, routing tasks to a tier by complexity; the lower tiers turn out to be indispensable rather than redundant, absorbing up to 24% of individual agent errors before they can compound into a patient\-facing mistake, and removing them causes the largest safety degradation of any ablation tested\[[41](https://arxiv.org/html/2608.06663#bib.bib1480)\]\. Ensemble Monitoring for AI Control finds that*diversity*among monitors, not the number of monitors or the compute spent on them, drives detection gains — a diverse three\-monitor ensemble beats a homogeneous one built from three copies of the same monitor by 2\.4x, even at equal compute — an oversight\-specific echo of the credit\-assignment lesson from §6 that how a signal is structured matters as much as how much of it there is\[[43](https://arxiv.org/html/2608.06663#bib.bib1483)\]\. “Managed Autonomy at Runtime” proposes a gear\-based framework that adjusts an agent’s operating autonomy level dynamically rather than fixing it in advance, explicitly for single\- and multi\-agent cyber\-operations settings where the appropriate oversight level is itself a function of context that changes during a long\-running task\[[77](https://arxiv.org/html/2608.06663#bib.bib1482)\]\. And a cautionary finding on the limits of any single mitigation, this time upstream at the training\-data layer rather than at runtime: fine\-tuning a model on synthetic agentic trajectories that include adversarial actions measurably increases misaligned behavior, and — the “phantom transfer” finding — this increase survives*removing every adversarial action from the training trajectories before fine\-tuning*, meaning the disposition toward misalignment was encoded diffusely across the whole trajectory rather than localized to the harmful steps a filter could catch\[[21](https://arxiv.org/html/2608.06663#bib.bib1491)\]— direct evidence that safety interventions aimed only at filtering visible harmful actions, whether at training time or at runtime, are an incomplete solution to a problem that originates upstream, in what the generating process teaches a model about its own disposition\.

Everything above is single\-episode diagnosis: a run derails, a constraint erodes, an oversight mechanism catches or misses one failure\. A separate and much smaller question is whether the capability to*avoid*these failures can itself accumulate across episodes, the way §3, §5, §6, and §7 each found a minority thread doing for their own topic\. EvoAgentBench takes this on directly for foundational measurement: it argues that what should transfer across a self\-evolving agent’s episodes is not information but*procedure*— reusable searching, debugging, and verification routines extracted from execution traces — and that this specific form of transfer is invisible to both single\-episode agent benchmarks and retention\-focused memory benchmarks, which is exactly why a dedicated instrument was needed\[[24](https://arxiv.org/html/2608.06663#bib.bib1460)\]\. This is the one point in this section where the cross\-task\-persistent locus of Axis 2 \(§2\.3\) is addressed directly rather than left for §9 to note as an absence; the fact that it took a benchmark paper to raise it, rather than a theoretical account, is itself consistent with this section’s Assessment\.

Assessment\.No general, empirically validated theory of long\-horizon degradation currently unifies this section’s findings; what exists is a set of well\-documented, partially contradictory phenomena \(bursty pre\-failure signals, bimodal rather than smooth coherence loss, drifting goals survivable by no single existing oversight mechanism\) that a future theory would need to explain simultaneously\. This is this survey’s most explicit limitation, not a hedge: the category’s small size relative to the corpus, even after a targeted supplement, is itself evidence that measurement and mitigation are ahead of theory here, and the field’s most useful near\-term contribution may be sharper diagnostics \(§7’s benchmarks, this section’s pre\-failure signals, EvoAgentBench’s procedural\-transfer measurement\) rather than a unifying account of why long horizons are hard\. We return to this gap, and to what the corpus’s own shape says about the field’s priorities, in the Discussion\.

## 9\. Discussion

The body sections each ended with an Assessment specific to their own literature\. Four threads recur across multiple sections and are worth naming explicitly, followed by an interpretive look at what the corpus’s own shape suggests about where the field is actually spending its effort — with the sampling caveats of §2\.2 stated again first, since every observation below is a hypothesis about the literature’s composition, not a finding about the underlying research problem’s difficulty\.

Shared evaluator assumptions across training and evaluation\.§6 and §7 converge on the same structural problem from opposite ends\. Training methods for long horizons need a reliable way to tell whether a trajectory is going well, and increasingly manufacture that signal via process reward models and credit\-assignment schemes that operate on intermediate steps rather than final outcomes\. Evaluation methods for long horizons need exactly the same thing — a reliable way to tell whether a trajectory is going well, independent of whether it was produced by a training method or an inference\-time harness — and increasingly manufacture it via trajectory\-level diagnostics and process\-aware benchmarks rather than terminal task success\. The implication worth flagging is that the tools used to validate whether long\-horizon training works \(§7’s process\-aware benchmarks\) and the tools used to build long\-horizon training in the first place \(§6’s process reward models\) rest on overlapping assumptions about what counts as progress on a partial trajectory\. We deliberately do not call this circularity: nothing we observed shows a specific benchmark inheriting a specific training signal, and independently constructed process signals can and do cross\-validate each other\. The narrower and better\-supported concern iscorrelated measurement bias— if the field’s shared intuitions about what “good intermediate progress” looks like are systematically off in some respect, that error would be partly invisible to a check that compares training methods against benchmarks resting on the same intuitions\. Establishing whether such a correlation exists, and how strong it is, would require deliberately constructing process signals from disjoint assumptions and comparing them — an experiment the corpus does not currently contain\. We flag this as a structural\-analogy hypothesis rather than a finding, and it is worth being explicit that it carries less evidentiary weight than the harness\-versus\-model question below: nothing in the corpus demonstrates this bias in a specific benchmark\-training pair, whereas the harness\-versus\-model question is backed by a concrete pattern in the corpus itself \(§5’s orchestration\-to\-recovery ratio, §3’s decomposition\-complexity finding\)\. We keep both in this Discussion because a structural risk worth naming does not have to be already measured to be worth naming — but the two should not be read as equally supported\.

Harness vs\. model: where does long\-horizon capability actually live?§3 and §5 both surface the same open attribution question from different angles: how much of what makes a system complete a longer task is the underlying model’s capability, and how much is the harness wrapped around it — better context management, better orchestration, better recovery logic — with a fixed model underneath? §5’s finding that orchestration and recovery engineering vastly outnumber loop\-level model\-capability papers in raw volume is suggestive but not dispositive; it could mean the harness genuinely carries most of the load, or simply that harness engineering is cheaper to iterate on than retraining a model, so the corpus reflects research cost structure rather than where capability actually lives\. §3’s finding that harnesses with more elaborate decomposition are not uniformly better complicates the “just build a better harness” reading further: past some point, additional scaffolding measurably hurts rather than helps, meaning harness quality itself has a capability ceiling that a fixed model cannot be scaffolded past\. Disentangling these two contributions — the model’s own long\-horizon competence versus the harness’s compensation for its absence — is, in our reading, the single most consequential open measurement problem this survey surfaces, because it bears directly on whether long\-horizon capability should be pursued by training better models or building better harnesses, and the corpus does not yet contain the controlled comparison that would answer it\.

Does long\-horizon reliability scale?§8 documents diagnostic evidence on both sides of whether degradation over long horizons follows any general, predictable curve — bursty pre\-failure signals and drifting goals suggesting correlated, worse\-than\-independent error on one side, high\-variance bimodal outcomes that resist a single smooth decay curve on the other — without resolving to either\. §7’s measured, reported \(not proven\) time\-horizon growth trend\[[45](https://arxiv.org/html/2608.06663#bib.bib1571)\]is the closest thing the field has to an aggregate answer, and it is explicitly a capability trend across model generations, not a reliability\-scaling law within a fixed model or harness\. Whether the field will find a predictive theory here, or whether “it depends on the failure mode” is the durable answer, remains open; we treat this as the survey’s most important flagged uncertainty rather than resolve it in either direction\.

The trajectory is becoming the unit of analysis\.A fourth pattern is visible only when the six sections are read together: across every category, the object being stored, scored, attributed over, audited, and debugged is shifting from the*outcome*to the*execution trace*\. In training, credit assignment operates over trajectory segments and trajectory graphs rather than final rewards \(§6\) — SALT assigns step\-level advantage by building a graph over the trajectory itself\[[50](https://arxiv.org/html/2608.06663#bib.bib570)\]\. In evaluation, Pass@1 is being displaced by trajectory\-level diagnosis, and the coherence\-collapse finding was only visible*because*the diagnostic decomposed 16,758 stored trajectories into aligned stages \(§7\)\[[40](https://arxiv.org/html/2608.06663#bib.bib933)\]\. In memory, the question has moved from what to store to what a stored trace can later be held accountable for: MemWeaver builds a memory framework explicitly for “traceable” long\-horizon reasoning with evidence\-grounded reuse\[[116](https://arxiv.org/html/2608.06663#bib.bib741)\], and Slipstream validates context compaction against the trajectory it compacts, treating the compactor’s unawareness of what will later be needed as a structural validation gap\[[16](https://arxiv.org/html/2608.06663#bib.bib1085)\]\. In safety, forensic trajectory signatures make memory poisoning detectable after the fact \(§8\)\[[48](https://arxiv.org/html/2608.06663#bib.bib1338)\]\. And in the self\-evolution literature that cuts across §§6\-8, EvoAgentBench argues the transferable unit is*procedural*— reusable search, debug, and verification procedures extracted from execution traces — which existing single\-episode agent benchmarks and information\-retention memory benchmarks both fail to isolate\[[24](https://arxiv.org/html/2608.06663#bib.bib1460)\]\. This convergence is worth naming because it has a practical consequence the individual sections do not state: trajectory retention is quietly becoming a precondition for progress on the other three threads above\. The harness\-versus\-model attribution question \(thread two\) is not answerable from outcome scores alone, and the correlated\-measurement\-bias concern \(thread one\) can only be tested by comparing process signals against each other at the step level\. The reliability\-scaling question \(thread three\) depends on it too, and in the most direct way of the three: §8’s own evidence that degradation is bursty and bimodal rather than smooth — the pre\-failure signal in Strained Coherence, the meltdown\-or\-profit split in Vending\-Bench — is itself trajectory\-level diagnosis, not an outcome statistic; a field that only logged pass/fail per episode would not have been able to observe either finding\. A field that discards its execution traces cannot run any of these three experiments — which makes trajectory logging, and the storage and privacy questions that come with it \(§4, §8\), infrastructure rather than an implementation detail\.

![Figure 6: Growth timeline, seed corpus only (n=1,419). Top: quarterly paper counts per category, log scale. Bottom: category share of quarterly output. The partial final quarter (2026Q3, 76 papers) is omitted from the trend and annotated separately.](https://arxiv.org/html/2608.06663v1/figures/growth_timeline.png)Figure 6: Growth timeline, seed corpus only \(n=1,419\)\. Top: quarterly paper counts per category, log scale\. Bottom: category share of quarterly output\. The partial final quarter \(2026Q3, 76 papers\) is omitted from the trend and annotated separately\.What the corpus’s shape suggests \(hypotheses, not findings\)\.Figure 6’s two panels showexecutionas the largest category by share throughout 2024\-2025 — consistently the plurality leader, fluctuating roughly 41\-54% quarter to quarter rather than holding a flat majority — andmemory’s share rising to match or slightly exceed it from 2026 Q1 onward, whileplanning’s share has drifted downward over the same period\. One explanation consistent with this pattern: the field solved a version of “make an agent take more than one action” \(execution/orchestration\) early and cheaply, and only once agents were routinely attempting genuinely long trajectories did the specific failure mode of losing track of earlier information \(memory\) become salient enough to attract comparable research volume — a herd\-dynamics\-style account in which research attention follows visible failure, not a claim that memory is intrinsically harder than execution\. Table 1’s 2026\-share column offers a second, complementary hypothesis:foundations\(77% posted in 2026\) andmemory\(72%\) skew far more recent thanplanning\(27%\), consistent with a verifiability\-gradient explanation — decomposition strategies are comparatively easy to demonstrate and publish quickly, while diagnosing why a system fails over a long horizon, or building memory architectures robust to interference at scale, requires the kind of large\-scale deployment experience that has only become available recently\. Both explanations are offered as testable hypotheses, in the sense that they predict specific further patterns \(e\.g\., a verifiability\-gradient account predicts foundations papers should lag their own triggering capability demonstrations by a roughly consistent delay\), not as conclusions this corpus alone can establish\.

## 10\. Conclusion

This survey organized 1,547 arXiv papers \(2024\-2026\) around a single question: as a task’s horizon grows past what one context window or one uninterrupted attempt can hold, what breaks first, and what compensates? Planning trades early commitment for robustness as environments grow less predictable \(§3\); memory trades fidelity for persistence across a three\-tier hierarchy of context, external storage, and weights, with no tier escaping that trade\-off \(§4\); execution control increasingly relies on the surrounding harness — orchestration and recovery — more than on the underlying model’s own reliability, though the two remain hard to disentangle \(§5\); training is moving from sparse outcome signals toward denser, step\-level credit as horizons lengthen, blurring the line between “outcome” and “process” supervision \(§6\); evaluation is undergoing the same shift, with a substantial and growing share of the literature devoted to showing that existing benchmarks do not measure what they claim to \(§7\); and the theory needed to unify these observations into a predictive account of long\-horizon degradation does not yet exist, though the diagnostic and safety literature building toward it is real and growing \(§8\)\. None of these six threads is complete on its own, and the Discussion’s cross\-cutting observations — the risk of correlated measurement bias between training and evaluation signals, the harness\-versus\-model attribution problem, whether long\-horizon reliability scales at all, and the field’s convergence on the execution trajectory as its shared unit of analysis — are, in our assessment, where the field’s next real progress has to be made\. We hope the taxonomy and the corpus behind it \(§ Data Availability\) make that progress easier to track\.

## Data Availability

The corpus \(corpus\_v2\.csv, 1,547 papers with taxonomy labels\), the classification and figure\-generation scripts, and the full bibliography \(references\.bib, auto\-generated;anchors\.bib, hand\-curated\) will be made available at a public repository accompanying this paper, to be linked in the camera\-ready/published version\. Classification rules, the harvest query threads, and the bleed\-quantification procedure described in §2\.2 are released as executable scripts rather than described only in prose, so that the corpus can be regenerated, audited, or extended\.

## References

## References

- \[1\]P\. Adamenko, M\. Ivanov, A\. Valeev, R\. Levichev, P\. Zadorozhny, I\. Lopatin, D\. Babayev, A\. Fenogenova, and V\. Malykh\(2025\)SWE\-MERA: a dynamic benchmark for agenticly evaluating large language models on software engineering tasks\.External Links:2507\.11059,[Link](https://arxiv.org/abs/2507.11059)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[2\]\(2025\)WOLF: werewolf\-based observations for LLM deception and falsehoods\.External Links:2512\.09187,[Link](https://arxiv.org/abs/2512.09187)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[3\]S\. Agashe, K\. Wong, V\. Tu, J\. Yang, A\. Li, and X\. E\. Wang\(2025\)Agent s2: a compositional generalist\-specialist framework for computer use agents\.External Links:2504\.00906,[Link](https://arxiv.org/abs/2504.00906)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[4\]M\. Ahn, A\. Brohan, N\. Brown, Y\. Chebotar, O\. Cortes, B\. David, C\. Finn, C\. Fu, K\. Gopalakrishnan, K\. Hausman, A\. Herzog, D\. Ho, J\. Hsu, J\. Ibarz, B\. Ichter, A\. Irpan, E\. Jang, R\. J\. Ruano, K\. Jeffrey, S\. Jesmonth, N\. J\. Joshi, R\. Julian, D\. Kalashnikov, Y\. Kuang, K\. Lee, S\. Levine, Y\. Lu, L\. Luu, C\. Parada, P\. Pastor, J\. Quiambao, K\. Rao, J\. Rettinghouse, D\. Reyes, P\. Sermanet, N\. Sievers, C\. Tan, A\. Toshev, V\. Vanhoucke, F\. Xia, T\. Xiao, P\. Xu, S\. Xu, M\. Yan, and A\. Zeng\(2022\)Do as i can, not as i say: grounding language in robotic affordances\.InConference on Robot Learning \(CoRL\),Proceedings of Machine Learning Research, Vol\.205,pp\. 287–318\.Note:SayCanExternal Links:2204\.01691,[Link](https://arxiv.org/abs/2204.01691)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[5\]R\. Aleithan, H\. Xue, M\. M\. Mohajer, E\. Nnorom, G\. Uddin, and S\. Wang\(2024\)SWE\-bench\+: enhanced coding benchmark for LLMs\.External Links:2410\.06992,[Link](https://arxiv.org/abs/2410.06992)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[6\]T\. An\(2025\)Cognitive workspace: active memory management for LLMs – an empirical study of functional infinite context\.External Links:2508\.13171,[Link](https://arxiv.org/abs/2508.13171)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[7\]R\. Arike, E\. Donoway, H\. Bartsch, and M\. Hobbhahn\(2025\)Technical report: evaluating goal drift in language model agents\.External Links:2505\.02709,[Link](https://arxiv.org/abs/2505.02709)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[8\]A\. Backlund and L\. Petersson\(2025\)Vending\-bench: a benchmark for long\-term coherence of autonomous agents\.External Links:2502\.15840,[Link](https://arxiv.org/abs/2502.15840)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p2.1)\.
- \[9\]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\.External Links:2508\.19005,[Link](https://arxiv.org/abs/2508.19005)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p6.1)\.
- \[10\]Y\. Cao, Y\. Zhong, Z\. Zeng, L\. Zheng, J\. Huang, H\. Qiu, P\. Shi, W\. Mao, and W\. Guanglu\(2026\)MobileDreamer: generative sketch world model for GUI agent\.External Links:2601\.04035,[Link](https://arxiv.org/abs/2601.04035)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p5.1)\.
- \[11\]F\. Carichon, A\. Khandelwal, M\. Fauchard, and G\. Farnadi\(2025\)The coming crisis of multi\-agent misalignment: AI alignment must be a dynamic and social process\.External Links:2506\.01080,[Link](https://arxiv.org/abs/2506.01080)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[12\]S\. Casper, L\. Bailey, R\. C\. Hunter, C\. Ezell, E\. Cabalé, M\. Gerovitch, S\. Slocum, K\. Wei, N\. Jurkovic, A\. Khan, P\. J\. K\. Christoffersen, A\. P\. Ozisik, R\. Trivedi, D\. Hadfield\-Menell, and N\. Kolt\(2025\)The AI agent index\.External Links:2502\.01635,[Link](https://arxiv.org/abs/2502.01635)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[13\]M\. Chen, L\. Wang, and B\. Qu\(2026\)Recursive self\-improvement in ai: from bounded self\-refinement to autonomous research loops\.Note:Companion survey; DeepGroundingExternal Links:2607\.07663,[Link](https://arxiv.org/abs/2607.07663)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p2.1),[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p4.1)\.
- \[14\]S\. Chen\(2026\)Governance decay: how context compaction silently erases safety constraints in long\-horizon LLM agents\.External Links:2606\.22528,[Link](https://arxiv.org/abs/2606.22528)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[15\]Y\. Chen, J\. Arkin, Y\. Hao, Y\. Zhang, N\. Roy, and C\. Fan\(2024\)PRompt optimization in multi\-step tasks \(PROMST\): integrating human feedback and heuristic\-based sampling\.External Links:2402\.08702,[Link](https://arxiv.org/abs/2402.08702)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p3.1)\.
- \[16\]Z\. Chen, R\. Pan, Y\. Dai, and R\. Netravali\(2026\)Slipstream: trajectory\-grounded compaction validation for long\-horizon agents\.External Links:2605\.08580,[Link](https://arxiv.org/abs/2605.08580)Cited by:[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[17\]X\. Cheng, S\. He, L\. Feng, H\. Xu, M\. Yan, L\. Feng, and B\. An\(2026\)Beyond trajectory\-level attribution: graph\-based credit assignment for agentic reinforcement learning\.External Links:2605\.26684,[Link](https://arxiv.org/abs/2605.26684)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[18\]P\. Dayan and G\. E\. Hinton\(1992\)Feudal reinforcement learning\.InAdvances in Neural Information Processing Systems 5 \(NIPS 1992\),External Links:[Link](https://proceedings.neurips.cc/paper/1992/hash/d14220ee66aeec73c49038385428ec4c-Abstract.html)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[19\]R\. F\. Del Rosario, K\. Krawiecka, and C\. S\. de Witt\(2025\)Architecting resilient LLM agents: a guide to secure plan\-then\-execute implementations\.External Links:2509\.08646,[Link](https://arxiv.org/abs/2509.08646)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p2.1)\.
- \[20\]X\. Deng, J\. Da, E\. Pan, Y\. He, C\. Ide, K\. Garg, N\. Lauffer, A\. Park, N\. Pasari, C\. Rane, K\. Sampath, M\. Krishnan, S\. Kundurthy, S\. M\. Hendryx, Z\. Wang, V\. Bharadwaj, J\. Holm, R\. Aluri, C\. B\. C\. Zhang, N\. Jacobson, B\. Liu, and B\. Kenstler\(2025\)SWE\-bench pro: can AI agents solve long\-horizon software engineering tasks?\.External Links:2509\.16941,[Link](https://arxiv.org/abs/2509.16941)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[21\]C\. Dixit\(2026\)Filtering harmful actions isn’t enough: phantom transfer in agentic SDF\.External Links:2607\.10750,[Link](https://arxiv.org/abs/2607.10750)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p4.1)\.
- \[22\]P\. Drammeh\(2025\)Multi\-agent LLM orchestration achieves deterministic, high\-quality decision support for incident response\.External Links:2511\.15755,[Link](https://arxiv.org/abs/2511.15755)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[23\]L\. E\. Erdogan, N\. Lee, S\. H\. Kim, S\. Moon, H\. Furuta, G\. K\. Anumanchipalli, K\. Keutzer, and A\. Gholami\(2025\)Plan\-and\-act: improving planning of agents for long\-horizon tasks\.External Links:2503\.09572,[Link](https://arxiv.org/abs/2503.09572)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[24\]X\. Gao, C\. Hu, H\. Chen, P\. Yao, Z\. Wang, Y\. Bai, Z\. Z\. Wu, Y\. Han, X\. Cong, J\. Gui, Y\. Deng, and T\. Li\(2026\)EvoAgentBench: benchmarking agent self\-evolution via ability transfer\.External Links:2607\.05202,[Link](https://arxiv.org/abs/2607.05202)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p6.1),[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p5.1),[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[25\]G\. Gonzalez\-Pumariega, S\. Agashe, J\. Yang, A\. Li, and X\. E\. Wang\(2026\)On the reliability of computer use agents\.External Links:2604\.17849,[Link](https://arxiv.org/abs/2604.17849)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[26\]Y\. Guo, L\. Xu, J\. Liu, D\. Ye, and S\. Qiu\(2025\)Segment policy optimization: effective segment\-level credit assignment in RL for large language models\.External Links:2505\.23564,[Link](https://arxiv.org/abs/2505.23564)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[27\]H\. Han, J\. Xie, X\. Ma, W\. Zhu, Z\. Zhang, Z\. Long, H\. Chen, and Q\. Ye\(2026\)SWE\-TRACE: optimizing long\-horizon SWE agents through rubric process reward models and heuristic test\-time scaling\.External Links:2604\.14820,[Link](https://arxiv.org/abs/2604.14820)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[28\]M\. M\. Hasan, H\. Li, E\. Fallahzadeh, G\. K\. Rajbahadur, B\. Adams, and A\. E\. Hassan\(2025\)An empirical study of testing practices in open source AI agent frameworks and agentic applications\.External Links:2509\.19185,[Link](https://arxiv.org/abs/2509.19185)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[29\]I\. Hong, C\. Yu, L\. Qiu, W\. Yan, Z\. Xu, H\. Jiang, Q\. Zhang, Q\. Lu, X\. Liu, C\. Zhang, and T\. Zhao\(2025\)Think\-RM: enabling long\-horizon reasoning in generative reward models\.External Links:2505\.16265,[Link](https://arxiv.org/abs/2505.16265)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[30\]J\. Huang and J\. T\. Zhou\(2026\)A two\-dimensional framework for AI agent design patterns: cognitive function and execution topology\.External Links:2605\.13850,[Link](https://arxiv.org/abs/2605.13850)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[31\]J\. Huang, X\. Chen, S\. Mishra, H\. S\. Zheng, A\. W\. Yu, X\. Song, and D\. Zhou\(2024\)Large language models cannot self\-correct reasoning yet\.Note:ICLR 2024External Links:2310\.01798,[Link](https://arxiv.org/abs/2310.01798)Cited by:[1\. Introduction](https://arxiv.org/html/2608.06663#Sx1.p5.1),[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[32\]W\. Huang, W\. Zhang, Y\. Liang, Y\. Bei, Y\. Chen, T\. Feng, X\. Pan, Z\. Tan, Y\. Wang, T\. Wei, S\. Wu, R\. Xu, L\. Yang, R\. Yang, W\. Yang, C\. Yeh, H\. Zhang, H\. Zhang, S\. Zhu, H\. P\. Zou, W\. Zhao, S\. Wang, W\. Xu, Z\. Ke, Z\. Hui, D\. Li, Y\. Wu, L\. He, C\. Wang, X\. Xu, B\. Huang, J\. Tan, S\. Heinecke, H\. Wang, C\. Xiong, A\. A\. Metwally, J\. Yan, C\. Lee, H\. Zeng, Y\. Xia, X\. Wei, A\. Payani, Y\. Wang, H\. Ma, W\. Wang, C\. Wang, Y\. Zhang, X\. Wang, Y\. Zhang, J\. You, H\. Tong, X\. Luo, X\. Liu, Y\. Sun, W\. Wang, J\. McAuley, J\. Zou, J\. Han, P\. S\. Yu, and K\. Shu\(2026\)Rethinking memory mechanisms of foundation agents in the second half: a survey\.External Links:2602\.06052,[Link](https://arxiv.org/abs/2602.06052)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[33\]X\. Huang, W\. Liu, X\. Chen, X\. Wang, H\. Wang, D\. Lian, Y\. Wang, R\. Tang, and E\. Chen\(2024\)Understanding the planning of LLM agents: a survey\.External Links:2402\.02716,[Link](https://arxiv.org/abs/2402.02716)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[34\]C\. E\. Jimenez, J\. Yang, A\. Wettig, S\. Yao, K\. Pei, O\. Press, and K\. Narasimhan\(2023\)SWE\-bench: can language models resolve real\-world GitHub issues?\.Note:ICLR 2024External Links:2310\.06770,[Link](https://arxiv.org/abs/2310.06770)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[35\]D\. Jin, W\. Sun, J\. Huang, P\. Liang, J\. Xuan, Y\. Liu, and Z\. Jin\(2025\)iReDev: a knowledge\-driven multi\-agent framework for intelligent requirements development\.External Links:2507\.13081,[Link](https://arxiv.org/abs/2507.13081)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[36\]T\. Joshi, S\. Chowdhury, and F\. Uysal\(2025\)SWE\-bench\-CL: continual learning for coding agents\.External Links:2507\.00014,[Link](https://arxiv.org/abs/2507.00014)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p6.1)\.
- \[37\]T\. Kagaya, S\. Lakshmi, A\. Ye, T\. J\. Yuan, J\. Karlekar, S\. Pranata, N\. Murakami, A\. Kinose, and Y\. You\(2025\)ViReSkill: vision\-grounded replanning with skill memory for LLM\-based planning in lifelong robot learning\.External Links:2509\.24219,[Link](https://arxiv.org/abs/2509.24219)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p5.1)\.
- \[38\]R\. Kamoi, Y\. Zhang, N\. Zhang, J\. Han, and R\. Zhang\(2024\)When can LLMs actually correct their own mistakes? a critical survey of self\-correction of LLMs\.External Links:2406\.01297,[Link](https://arxiv.org/abs/2406.01297)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[39\]M\. Kang, W\. Chen, D\. Han, H\. A\. Inan, L\. Wutschitz, Y\. Chen, R\. B\. Sim, and S\. Rajmohan\(2025\)ACON: optimizing context compression for long\-horizon LLM agents\.External Links:2510\.00615,[Link](https://arxiv.org/abs/2510.00615)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[40\]M\. Kim, D\. Wang, S\. Cui, F\. Farmahinifarahani, T\. Y\. Zhuo, S\. Garg, B\. Ray, R\. Mukherjee, and V\. Kumar\(2026\)Coherence collapse: diagnosing why code agents fail after reaching the right code\.External Links:2603\.24631,[Link](https://arxiv.org/abs/2603.24631)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1),[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p2.1),[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[41\]Y\. Kim, H\. Jeong, C\. Park, E\. Park, H\. Zhang, X\. Liu, H\. Lee, D\. McDuff, M\. Ghassemi, C\. Breazeal, S\. Tulebaev, and H\. W\. Park\(2025\)Tiered agentic oversight: a hierarchical multi\-agent system for healthcare safety\.External Links:2506\.12482,[Link](https://arxiv.org/abs/2506.12482)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p4.1)\.
- \[42\]N\. Kolt\(2025\)Governing AI agents\.External Links:2501\.07913,[Link](https://arxiv.org/abs/2501.07913)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[43\]E\. Koran, Y\. Yun, S\. Tetef, B\. Arnav, and P\. Bernabeu\-Pérez\(2026\)Ensemble monitoring for AI control: diverse signals outweigh more compute\.External Links:2605\.15377,[Link](https://arxiv.org/abs/2605.15377)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p4.1)\.
- \[44\]T\. Kuntz, A\. Duzan, H\. Zhao, F\. Croce, J\. Z\. Kolter, N\. Flammarion, and M\. Andriushchenko\(2025\)OS\-harm: a benchmark for measuring safety of computer use agents\.External Links:2506\.14866,[Link](https://arxiv.org/abs/2506.14866)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[45\]T\. Kwa, B\. West, J\. Becker, A\. Deng, K\. Garcia, M\. Hasin, S\. Jawhar, M\. Kinniment, N\. Rush, S\. V\. Arx, R\. Bloom, T\. Broadley, H\. Du, B\. Goodrich, N\. Jurkovic, L\. H\. Miles, S\. Nix, T\. Lin, C\. Painter, N\. Parikh, D\. Rein, L\. J\. K\. Sato, H\. Wijk, D\. M\. Ziegler, E\. Barnes, and L\. Chan\(2025\)Measuring AI ability to complete long software tasks\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:METR\. Originally titled "Measuring AI Ability to Complete Long Tasks" \(arXiv v1–v2, Mar 2025\); retitled to add "Software" from v3 onward\. NeurIPS 2025 main track camera\-ready author list omits Chris Painter, present on arXiv; author list here follows the arXiv versionExternal Links:2503\.14499,[Link](https://arxiv.org/abs/2503.14499)Cited by:[1\. Introduction](https://arxiv.org/html/2608.06663#Sx1.p1.1),[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p4.1),[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p4.1)\.
- \[46\]E\. J\. Laird and C\. Clark\(2025\)On memory: a comparison of memory mechanisms in world models\.External Links:2512\.06983,[Link](https://arxiv.org/abs/2512.06983)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p5.1)\.
- \[47\]H\. Lee, J\. C\. Chen, J\. Singh, Z\. Khan, E\. Stengel\-Eskin, and M\. Bansal\(2026\)MINTEval: evaluating memory under multi\-target interference in long\-horizon agent systems\.External Links:2605\.18565,[Link](https://arxiv.org/abs/2605.18565)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[48\]J\. W\. Leong\(2026\)Forensic trajectory signatures for agent memory poisoning detection\.External Links:2606\.30566,[Link](https://arxiv.org/abs/2606.30566)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p5.1),[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[49\]H\. F\. Leung, X\. Xi, and F\. Zuo\(2025\)AndroidControl\-curated: revealing the true potential of GUI agents through benchmark purification\.External Links:2510\.18488,[Link](https://arxiv.org/abs/2510.18488)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[50\]J\. Li, Y\. Wang, D\. Yan, Y\. Tian, Z\. Xu, H\. Song, P\. Xu, and L\. L\. Cheong\(2025\)SALT: step\-level advantage assignment for long\-horizon agents via trajectory graph\.External Links:2510\.20022,[Link](https://arxiv.org/abs/2510.20022)Cited by:[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[51\]J\. Li, Y\. Li, C\. Zhao, Z\. Xu, B\. Hu, and M\. Z\. \(111999\)\(2026\)WindowsWorld: a process\-centric benchmark of autonomous GUI agents in professional cross\-application environments\.External Links:2604\.27776,[Link](https://arxiv.org/abs/2604.27776)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[52\]R\. Li, X\. Zhang, H\. Yu, S\. Duan, X\. Li, W\. Xiang, C\. Liao, X\. Guo, Y\. Li, and J\. Suo\(2026\)MemPO: self\-memory policy optimization for long\-horizon agents\.External Links:2603\.00680,[Link](https://arxiv.org/abs/2603.00680)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p4.1)\.
- \[53\]X\. Li, H\. Li, Y\. Zhou, Y\. Liu, and Q\. Ai\(2026\)Beyond experience retrieval: learning to generate utility\-optimized structured experience for frozen LLMs\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2602.02556),2602\.02556,[Link](https://arxiv.org/abs/2602.02556)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p6.1)\.
- \[54\]Y\. Li\(2025\)Decomposing LLM self\-correction: the accuracy\-correction paradox and error depth hypothesis\.External Links:2601\.00828,[Link](https://arxiv.org/abs/2601.00828)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[55\]Y\. Li, B\. Xu, X\. Tian, X\. Xu, and H\. Shen\(2026\)Beyond entangled planning: task\-decoupled planning for long\-horizon agents\.External Links:2601\.07577,[Link](https://arxiv.org/abs/2601.07577)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[56\]S\. Liang, S\. Garg, and R\. Z\. Moghaddam\(2025\)The SWE\-bench illusion: when state\-of\-the\-art LLMs remember instead of reason\.External Links:2506\.12286,[Link](https://arxiv.org/abs/2506.12286)Cited by:[1\. Introduction](https://arxiv.org/html/2608.06663#Sx1.p5.1),[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[57\]H\. Lightman, V\. Kosaraju, Y\. Burda, H\. Edwards, B\. Baker, T\. Lee, J\. Leike, J\. Schulman, I\. Sutskever, and K\. Cobbe\(2023\)Let’s verify step by step\.Note:ICLR 2024External Links:2305\.20050,[Link](https://arxiv.org/abs/2305.20050)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[58\]J\. W\. Lin, E\. K\. Jones, D\. J\. Jasper, E\. Ho, A\. H\. Wu, A\. T\. Yang, N\. Perry, A\. Zou, M\. Fredrikson, J\. Z\. Kolter, P\. Liang, D\. Boneh, and D\. E\. Ho\(2025\)Comparing AI agents to cybersecurity professionals in real\-world penetration testing\.External Links:2512\.09882,[Link](https://arxiv.org/abs/2512.09882)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[59\]B\. Liu, Y\. Jiang, X\. Zhang, Q\. Liu, S\. Zhang, J\. Biswas, and P\. Stone\(2023\)LLM\+P: empowering large language models with optimal planning proficiency\.Note:arXiv preprint onlyExternal Links:2304\.11477,[Link](https://arxiv.org/abs/2304.11477)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[60\]G\. Liu, S\. Geng, S\. Li, H\. Cui, S\. Zhang, X\. Liu, and T\. Liu\(2025\)WebCoach: self\-evolving web agents with cross\-session memory guidance\.External Links:2511\.12997,[Link](https://arxiv.org/abs/2511.12997)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[61\]N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang\(2024\)Lost in the middle: how language models use long contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.Note:arXiv:2307\.03172, posted 2023External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[62\]X\. Liu, H\. Yu, H\. Zhang, Y\. Xu, X\. Lei, H\. Lai, Y\. Gu, H\. Ding, K\. Men, K\. Yang, S\. Zhang, X\. Deng, A\. Zeng, Z\. Du, C\. Zhang, S\. Shen, T\. Zhang, Y\. Su, H\. Sun, M\. Huang, Y\. Dong, and J\. Tang\(2023\)AgentBench: evaluating LLMs as agents\.Note:ICLR 2024External Links:2308\.03688,[Link](https://arxiv.org/abs/2308.03688)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[63\]P\. D\. Luong, L\. N\. Bao, N\. V\. K\. Tam, D\. H\. N\. Khoa, N\. H\. Quyen, V\. Pham, and P\. T\. Duy\(2025\)XOffense: an autonomous multi\-agent framework for penetration testing with domain\-adapted large language models\.External Links:2509\.13021,[Link](https://arxiv.org/abs/2509.13021)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[64\]Q\. Mao, Q\. Liu, Z\. Li, M\. Cheng, Z\. Zhang, and R\. Li\(2024\)PoTable: towards systematic thinking via plan\-then\-execute stage reasoning on tables\.External Links:2412\.04272,[Link](https://arxiv.org/abs/2412.04272)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[65\]M\. Martínez and X\. Franch\(2025\)Dissecting the SWE\-bench leaderboards: profiling submitters and architectures of LLM\- and agent\-based repair systems\.External Links:2506\.17208,[Link](https://arxiv.org/abs/2506.17208)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[66\]M\. Martínez and X\. Franch\(2026\)What’s in a benchmark? the case of SWE\-bench in automated program repair\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2602.04449),2602\.04449,[Link](https://arxiv.org/abs/2602.04449)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[67\]A\. Mehta and A\. Datta\(2026\)Plans don’t persist: why context management is load bearing for LLM agents\.External Links:2606\.22953,[Link](https://arxiv.org/abs/2606.22953)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[68\]G\. Mialon, C\. Fourrier, C\. Swift, T\. Wolf, Y\. LeCun, and T\. Scialom\(2023\)GAIA: a benchmark for general AI assistants\.Note:ICLR 2024External Links:2311\.12983,[Link](https://arxiv.org/abs/2311.12983)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[69\]C\. Mou, Z\. Zhuang, X\. Chen, and Y\. Zhang\(2026\)Beyond uniform credit assignment: selective eligibility traces for RLVR\.External Links:2605\.05965,[Link](https://arxiv.org/abs/2605.05965)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[70\]D\. D\. Nguyen, J\. Chen, Y\. Wang, G\. Wu, N\. Park, Z\. Hu, H\. Lyu, J\. Wu, R\. Aponte, Y\. Xia, X\. Li, J\. Shi, H\. Chen, V\. D\. Lai, Z\. Xie, S\. Kim, R\. Zhang, Y\. Tong, M\. Tanjim, N\. K\. Ahmed, P\. Mathur, S\. Yoon, L\. Yao, B\. Kveton, J\. Kil, T\. H\. Nguyen, T\. Bui, T\. Zhou, R\. A\. Rossi, and F\. Dernoncourt\(2024\)GUI agents: a survey\.External Links:2412\.13501,[Link](https://arxiv.org/abs/2412.13501)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[71\]Z\. Ni, Y\. Li, and D\. Dong\(2024\)Tree\-of\-code: a hybrid approach for robust complex task planning and execution\.External Links:2412\.14212,[Link](https://arxiv.org/abs/2412.14212)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[72\]N\. Otani, N\. Bhutani, H\. Kim, D\. Zhang, and E\. Hruschka\(2026\)Do agents need to plan step\-by\-step? rethinking planning horizon in data\-centric tool calling\.External Links:2605\.08477,[Link](https://arxiv.org/abs/2605.08477)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p3.1)\.
- \[73\]C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez\(2023\)MemGPT: towards LLMs as operating systems\.Note:arXiv preprint; not a peer\-reviewed venue \(despite being widely but incorrectly cited around the web as COLM 2024\)External Links:2310\.08560,[Link](https://arxiv.org/abs/2310.08560)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[74\]M\. Pandya, K\. Zhang, and B\. Lyu\(2026\)Strained coherence: a pre\-failure signal in coding agent execution trajectories\.External Links:2606\.07889,[Link](https://arxiv.org/abs/2606.07889)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p2.1)\.
- \[75\]J\. S\. Park, J\. C\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein\(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology \(UIST\),External Links:[Document](https://dx.doi.org/10.1145/3586183.3606763),2304\.03442,[Link](https://arxiv.org/abs/2304.03442)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[76\]T\. Prathifkumar, N\. S\. Mathews, and M\. Nagappan\(2025\)Does SWE\-bench\-verified test agent ability or model memory?\.External Links:2512\.10218,[Link](https://arxiv.org/abs/2512.10218)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[77\]S\. Ramaswamy and W\. Miaosheng\(2026\)Managed autonomy at runtime: gear\-based safety and governance for single\- and multi\-agent cyber\-physical systems\.External Links:2607\.00334,[Link](https://arxiv.org/abs/2607.00334)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p4.1)\.
- \[78\]T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom\(2023\)Toolformer: language models can teach themselves to use tools\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:Oral\. Author list follows the NeurIPS camera\-ready, which adds Eric Hambro relative to the single\-version arXiv preprint \(8 authors, no Hambro\)External Links:2302\.04761,[Link](https://arxiv.org/abs/2302.04761)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[79\]J\. Shao, Y\. Lu, Y\. Zhang, and B\. Luo\(2026\)When stored evidence stops being usable: scale\-conditioned evaluation of agent memory\.External Links:2605\.07313,[Link](https://arxiv.org/abs/2605.07313)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[80\]Y\. Shen, K\. Song, X\. Tan, D\. Li, W\. Lu, and Y\. Zhuang\(2023\)HuggingGPT: solving AI tasks with ChatGPT and its friends in hugging face\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:2303\.17580,[Link](https://arxiv.org/abs/2303.17580)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[81\]L\. Sheng and S\. Xu\(2025\)CSC\-SQL: corrective self\-consistency in text\-to\-SQL via reinforcement learning\.External Links:2505\.13271,[Link](https://arxiv.org/abs/2505.13271)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[82\]J\. Shi, T\. J\. Zhang, Z\. Jin, and V\. Conitzer\(2026\)Cheap talk, empty promise: frontier LLMs easily break public promises for self\-interest\.External Links:2604\.04782,[Link](https://arxiv.org/abs/2604.04782)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[83\]N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. Yao\(2023\)Reflexion: language agents with verbal reinforcement learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:2303\.11366,[Link](https://arxiv.org/abs/2303.11366)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[84\]S\. Silver, J\. Sun, I\. Zhang, S\. Hooker, and E\. Kim\(2025\)Language models can perform single\-utterance self\-correction of perturbed reasoning\.External Links:2506\.15894,[Link](https://arxiv.org/abs/2506.15894)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[85\]M\. Sullivan and A\. Koller\(2025\)GRPO is secretly a process reward model\.External Links:2509\.21154,[Link](https://arxiv.org/abs/2509.21154)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[86\]W\. Sun, M\. Lu, Z\. Ling, K\. Liu, X\. Yao, Y\. Yang, and J\. Chen\(2025\)Scaling long\-horizon LLM agent via context\-folding\.External Links:2510\.11967,[Link](https://arxiv.org/abs/2510.11967)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[87\]Z\. Sun, Z\. Liu, Y\. Zang, Y\. Cao, X\. Dong, T\. Wu, D\. Lin, and J\. Wang\(2025\)SEAgent: self\-evolving computer use agent with autonomous learning from experience\.External Links:2508\.04700,[Link](https://arxiv.org/abs/2508.04700)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p5.1)\.
- \[88\]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\.Note:The Options FrameworkExternal Links:[Document](https://dx.doi.org/10.1016/S0004-3702%2899%2900052-1)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[89\]F\. Tang, H\. Xu, H\. Zhang, S\. Chen, X\. Wu, Y\. Shen, W\. Zhang, G\. Hou, Z\. Tan, Y\. Yan, K\. Song, J\. Shao, W\. Lü, J\. Xiao, and Y\. Zhuang\(2025\)A survey on \(m\)LLM\-based GUI agents\.External Links:2504\.13865,[Link](https://arxiv.org/abs/2504.13865)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[90\]M\. Tantakoun, X\. Zhu, and C\. Muise\(2025\)LLMs as planning formalizers: a survey for leveraging large language models to construct automated planning models\.External Links:2503\.18971,[Link](https://arxiv.org/abs/2503.18971)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[91\]B\. Wang, B\. Li, M\. Wang, Y\. Tao, and F\. Kong\(2026\)Harnesses for inference\-time alignment over execution trajectories\.External Links:2605\.21516,[Link](https://arxiv.org/abs/2605.21516)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[92\]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\.Note:TMLR 2024External Links:2305\.16291,[Link](https://arxiv.org/abs/2305.16291)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p5.1)\.
- \[93\]H\. Wang, S\. Yang, Y\. Chen, and P\. Liu\(2026\)Hijacking agent memory: stealthy trojan attacks through conversational interaction\.External Links:2605\.29960,[Link](https://arxiv.org/abs/2605.29960)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p5.1)\.
- \[94\]Y\. Wang, M\. Pradel, and Z\. Liu\(2025\)Are "solved issues" in SWE\-bench really solved correctly? an empirical study\.External Links:2503\.15223,[Link](https://arxiv.org/abs/2503.15223)Cited by:[1\. Introduction](https://arxiv.org/html/2608.06663#Sx1.p5.1),[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[95\]L\. Wei, X\. Peng, X\. Dong, N\. Xie, and B\. Wang\(2026\)FadeMem: biologically\-inspired forgetting for efficient agent memory\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2601.18642),2601\.18642,[Link](https://arxiv.org/abs/2601.18642)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[96\]J\. Wu, M\. Hu, J\. Zhu, J\. Pan, Y\. Liu, M\. Xu, and Y\. Jin\(2025\)Git context controller: manage the context of LLM\-based agents like git\.External Links:2508\.00031,[Link](https://arxiv.org/abs/2508.00031)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[97\]M\. Wu, Y\. Wang, Y\. Ming, Y\. An, Y\. Wan, W\. Chen, B\. Lin, Y\. Li, T\. Xie, and D\. Zhou\(2025\)CheMatAgent: enhancing LLMs for chemistry and materials science through tree\-search based tool learning\.External Links:2506\.07551,[Link](https://arxiv.org/abs/2506.07551)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p4.1)\.
- \[98\]M\. Wu, D\. Zhang, Y\. Li, D\. Zhou, and W\. Chen\(2025\)SELT: self\-evaluation tree search for LLMs with task decomposition\.External Links:2506\.07557,[Link](https://arxiv.org/abs/2506.07557)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p4.1)\.
- \[99\]X\. Wu, Q\. Sun, R\. Zhang, C\. Song, J\. Wu, Y\. Qi, and H\. Cheng\(2026\)Demystifying reinforcement learning for long\-horizon tool\-using agents: a comprehensive recipe\.External Links:2603\.21972,[Link](https://arxiv.org/abs/2603.21972)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[100\]X\. Wu, Z\. Li, G\. Shi, A\. Duffy, T\. Marques, M\. Olson, T\. Zhou, and D\. Manocha\(2026\)Co\-evolving LLM decision and skill bank agents for long\-horizon tasks\.External Links:2604\.20987,[Link](https://arxiv.org/abs/2604.20987)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p4.1)\.
- \[101\]Z\. Wu, Q\. Zeng, Z\. Zhang, Z\. Tan, C\. Shen, and M\. Jiang\(2024\)Large language models can self\-correct with key condition verification\.External Links:2405\.14092,[Link](https://arxiv.org/abs/2405.14092)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[102\]Z\. Xi, C\. Liao, G\. Li, Y\. Yang, W\. Chen, Z\. Zhang, B\. Wang, S\. Jin, Y\. Zhou, J\. Guan, W\. Wu, T\. Ji, T\. Gui, Q\. Zhang, and X\. Huang\(2025\)AgentPRM: process reward models for LLM agents via step\-wise promise and progress\.External Links:2511\.08325,[Link](https://arxiv.org/abs/2511.08325)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[103\]S\. Xia, Y\. Wang, Z\. Z\. Huang, and P\. Liu\(2026\)Diagnosing and mitigating context rot in long\-horizon search\.External Links:2606\.29718,[Link](https://arxiv.org/abs/2606.29718)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p2.1)\.
- \[104\]T\. Xie, D\. Zhang, J\. Chen, X\. Li, S\. Zhao, R\. Cao, T\. J\. Hua, Z\. Cheng, D\. Shin, F\. Lei, Y\. Liu, Y\. Xu, S\. Zhou, S\. Savarese, C\. Xiong, V\. Zhong, and T\. Yu\(2024\)OSWorld: benchmarking multimodal agents for open\-ended tasks in real computer environments\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:2404\.07972,[Link](https://arxiv.org/abs/2404.07972)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[105\]B\. Xu, T\. Zheng, Z\. Wang, H\. T\. Tsang, W\. Wang, T\. Fang, and Y\. Song\(2025\)The cognitive bandwidth bottleneck: shifting long\-horizon agent from planning with actions to planning with schemas\.External Links:2510\.07091,[Link](https://arxiv.org/abs/2510.07091)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p3.1)\.
- \[106\]W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. Zhang\(2025\)A\-MEM: agentic memory for LLM agents\.External Links:2502\.12110,[Link](https://arxiv.org/abs/2502.12110)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[107\]X\. Xu, Z\. Zhao, C\. Song, Z\. Wang, J\. Qiang, J\. Yan, and Y\. Lin\(2025\)SentiMM: a multimodal multi\-agent framework for sentiment analysis in social media\.External Links:2508\.18108,[Link](https://arxiv.org/abs/2508.18108)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[108\]S\. Yan, A\. Bahloul, E\. Nie, S\. Schwarzmann, R\. Trivisonno, V\. Tresp, and Y\. Ma\(2026\)Memory\-r2: fair credit assignment for long\-horizon memory\-augmented LLM agents\.External Links:2605\.21768,[Link](https://arxiv.org/abs/2605.21768)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[109\]C\. Yang, C\. Zhou, Y\. Xiao, S\. Dong, L\. Zhuang, Y\. Zhang, Z\. Wang, Z\. Hong, Z\. Yuan, Z\. Xiang, S\. Chen, H\. Zhou, Q\. Zhang, N\. Liu, J\. Su, B\. An, Y\. Chang, and X\. Huang\(2026\)Graph\-based agent memory: taxonomy, techniques, and applications\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2602.05665),2602\.05665,[Link](https://arxiv.org/abs/2602.05665)Cited by:[4\. Memory & Context Management](https://arxiv.org/html/2608.06663#Sx4.p3.1)\.
- \[110\]J\. Yang, C\. Jimenez\-Gomez, A\. Zhang, K\. Lieret, J\. Yang, X\. Wu, O\. Press, N\. Muennighoff, G\. Synnaeve, K\. Narasimhan, D\. Yang, S\. I\. Wang, and O\. Press\(2024\)SWE\-bench multimodal: do AI systems generalize to visual software domains?\.External Links:2410\.03859,[Link](https://arxiv.org/abs/2410.03859)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[111\]P\. Yang, H\. Ci, and M\. Z\. Shou\(2025\)macOSWorld: a multilingual interactive benchmark for GUI agents\.External Links:2506\.04135,[Link](https://arxiv.org/abs/2506.04135)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[112\]Y\. Yang, S\. Padarha, S\. Ball, A\. Lee, and A\. Mahdi\(2025\)Agentic reinforcement learning for search misaligns instruction\-tuning\.External Links:2510\.17431,[Link](https://arxiv.org/abs/2510.17431)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[113\]S\. Yao, N\. Shinn, P\. Razavi, and K\. Narasimhan\(2024\)τ\\tau\-Bench: a benchmark for tool\-agent\-user interaction in real\-world domains\.Note:ICLR 2025External Links:2406\.12045,[Link](https://arxiv.org/abs/2406.12045)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[114\]S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. L\. Griffiths, Y\. Cao, and K\. Narasimhan\(2023\)Tree of thoughts: deliberate problem solving with large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:2305\.10601,[Link](https://arxiv.org/abs/2305.10601)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p4.1)\.
- \[115\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2022\)ReAct: synergizing reasoning and acting in language models\.Note:ICLR 2023External Links:2210\.03629,[Link](https://arxiv.org/abs/2210.03629)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p2.1)\.
- \[116\]J\. Ye, X\. Li, X\. Yang, C\. Huang, L\. Nie, L\. Yao, and D\. Zhan\(2026\)MemWeaver: weaving hybrid memories for traceable long\-horizon agentic reasoning\.External Links:2601\.18204,[Link](https://arxiv.org/abs/2601.18204)Cited by:[9\. Discussion](https://arxiv.org/html/2608.06663#Sx9.p5.1)\.
- \[117\]W\. Ye, B\. Yuan, Z\. Xu, Y\. Tian, Y\. Wang, H\. Kautz, and A\. Zhang\(2026\)A sober look at agentic misalignment in automated workflows\.External Links:2605\.24197,[Link](https://arxiv.org/abs/2605.24197)Cited by:[8\. Foundations, Limits & Safety](https://arxiv.org/html/2608.06663#Sx8.p3.1)\.
- \[118\]D\. Zan, Z\. Huang, W\. Liu, H\. Chen, L\. Zhang, S\. Xin, C\. Lü, Q\. Liu, X\. Zhong, A\. Li, S\. Liu, Y\. Xiao, L\. Chen, Y\. Zhang, J\. Su, T\. Liu, R\. Long, K\. Shen, and X\. Liang\(2025\)Multi\-SWE\-bench: a multilingual benchmark for issue resolving\.External Links:2504\.02605,[Link](https://arxiv.org/abs/2504.02605)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[119\]D\. Zan, Z\. Huang, A\. Yu, S\. Lin, Y\. Shi, W\. Liu, D\. Chen, Z\. Qi, H\. Yu, L\. Yu, D\. Ran, M\. Zeng, B\. Shen, P\. Bian, G\. Liang, B\. Guan, P\. Huang, T\. Xie, Y\. Wang, and Q\. Wang\(2024\)SWE\-bench\-java: a GitHub issue resolving benchmark for java\.External Links:2408\.14354,[Link](https://arxiv.org/abs/2408.14354)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1)\.
- \[120\]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\.External Links:2511\.10395,[Link](https://arxiv.org/abs/2511.10395)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p6.1)\.
- \[121\]C\. Zhang, S\. He, J\. Qian, B\. Li, L\. Li, S\. Qin, K\. Yu, M\. Ma, G\. M\. Liu, Q\. Lin, S\. Rajmohan, D\. Zhang, and Q\. Zhang\(2024\)Large language model\-brained GUI agents: a survey\.External Links:2411\.18279,[Link](https://arxiv.org/abs/2411.18279)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[122\]G\. Zhang, H\. Geng, X\. Yu, Z\. Yin, Z\. Zhang, Z\. Tan, H\. Zhou, Z\. Li, X\. Xue, Y\. Li, Y\. Zhou, Y\. Chen, C\. Zhang, Y\. Fan, Z\. Wang, S\. Huang, F\. Piedrahita\-Velez, Y\. Liao, H\. Wang, M\. Yang, H\. Ji, J\. Wang, S\. Yan, P\. H\. S\. Torr, and L\. Bai\(2025\)The landscape of agentic reinforcement learning for LLMs: a survey\.External Links:2509\.02547,[Link](https://arxiv.org/abs/2509.02547)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[123\]H\. Zhang, P\. Wang, S\. Diao, Y\. Lin, R\. Pan, H\. Dong, D\. Zhang, P\. Molchanov, and T\. Zhang\(2024\)Entropy\-regularized process reward model\.External Links:2412\.11006,[Link](https://arxiv.org/abs/2412.11006)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[124\]L\. Zhang, Y\. Xiao, X\. Lu, J\. Cao, Y\. Zhao, J\. Zhou, L\. An, Z\. Feng, W\. Sha, Y\. Shi, C\. Xiao, J\. Xiong, Y\. Zhang, H\. Wu, and H\. Wang\(2026\)OmegaUse: building a general\-purpose GUI agent for autonomous task execution\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2601.20380),2601\.20380,[Link](https://arxiv.org/abs/2601.20380)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[125\]N\. Zhang, R\. Sun, J\. Su, H\. Yang, Z\. Pan, Z\. Chen, and X\. Zhang\(2026\)MICA: multi\-granularity intertemporal credit assignment for long\-horizon emotional support dialogue\.Open MIND\.External Links:[Document](https://dx.doi.org/10.48550/arxiv.2603.06194),2603\.06194,[Link](https://arxiv.org/abs/2603.06194)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p2.1)\.
- \[126\]T\. Zhang, A\. Popa, Y\. Xu, R\. Song, and D\. Dimitriadis\(2026\)PIVOT: bridging planning and execution in LLM agents via trajectory refinement\.External Links:2605\.11225,[Link](https://arxiv.org/abs/2605.11225)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p2.1)\.
- \[127\]Y\. Zhang, Y\. Song, S\. Li, Y\. Shi, S\. Shen, X\. Tang, M\. Zhang, and D\. Wang\(2025\)Generative AI\-driven hierarchical multi\-agent framework for zero\-touch optical networks\.External Links:2510\.05625,[Link](https://arxiv.org/abs/2510.05625)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p3.1)\.
- \[128\]Y\. Zhang, Y\. Wu, H\. Zhang, W\. Li, H\. Chen, J\. Wu, G\. Li, Z\. Han, and V\. Tresp\(2025\)GroundedPRM: tree\-guided and fidelity\-aware process reward modeling for step\-level reasoning\.External Links:2510\.14942,[Link](https://arxiv.org/abs/2510.14942)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p3.1)\.
- \[129\]Y\. Zhang, M\. Fang, Z\. Chen, and M\. Pechenizkiy\(2026\)Self\-evolving LLM agents with in\-distribution optimization\.External Links:2606\.07367,[Link](https://arxiv.org/abs/2606.07367)Cited by:[6\. Training for Long Horizons](https://arxiv.org/html/2608.06663#Sx6.p4.1)\.
- \[130\]Z\. Zhang, X\. Bo, C\. Ma, R\. Li, X\. Chen, Q\. Dai, J\. Zhu, Z\. Dong, and J\. Wen\(2024\)A survey on the memory mechanism of large language model based agents\.External Links:2404\.13501,[Link](https://arxiv.org/abs/2404.13501)Cited by:[2\.4 Positioning vs\. existing surveys](https://arxiv.org/html/2608.06663#Sx2.SSx4.p1.1)\.
- \[131\]X\. Zhao, F\. Gao, X\. Song, Y\. Chen, R\. Yang, Y\. Fu, Y\. Wang, Y\. Iwasawa, Y\. Matsuo, and I\. Li\(2025\)ReAgent: reversible multi\-agent reasoning for knowledge\-enhanced multi\-hop QA\.External Links:2503\.06951,[Link](https://arxiv.org/abs/2503.06951)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p3.1)\.
- \[132\]J\. Zhou, J\. Chen, Q\. Lu, D\. Zhao, and L\. Zhu\(2025\)SHIELDA: structured handling of exceptions in LLM\-driven agentic workflows\.External Links:2508\.07935,[Link](https://arxiv.org/abs/2508.07935)Cited by:[5\. Execution Control & Recovery](https://arxiv.org/html/2608.06663#Sx5.p4.1)\.
- \[133\]S\. Zhou, F\. Yue, Z\. Hu, Y\. Shen, C\. Zhang, feng hong, C\. Liu, and K\. Zeng\(2026\)ToolVerse: unlocking massive environments and long\-horizon tasks for agentic reinforcement learning\.External Links:2607\.15660,[Link](https://arxiv.org/abs/2607.15660)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p5.1)\.
- \[134\]S\. Zhou, F\. F\. Xu, H\. Zhu, X\. Zhou, R\. Lo, A\. Sridhar, X\. Cheng, T\. Ou, Y\. Bisk, D\. Fried, U\. Alon, and G\. Neubig\(2023\)WebArena: a realistic web environment for building autonomous agents\.Note:ICLR 2024External Links:2307\.13854,[Link](https://arxiv.org/abs/2307.13854)Cited by:[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p3.1)\.
- \[135\]S\. Zhou, T\. Zhou, Y\. Yang, G\. Long, D\. Ye, J\. Jiang, and C\. Zhang\(2024\)WALL\-e: world alignment by rule learning improves world model\-based LLM agents\.External Links:2410\.07484,[Link](https://arxiv.org/abs/2410.07484)Cited by:[3\. Planning & Decomposition](https://arxiv.org/html/2608.06663#Sx3.p5.1)\.
- \[136\]Y\. Zhu, T\. Jin, Y\. Pruksachatkun, A\. Zhang, S\. Liu, S\. Cui, S\. Kapoor, S\. Longpre, K\. Meng, R\. Weiss, F\. Barez, R\. Gupta, J\. Dhamala, J\. Merizian, M\. Giulianelli, H\. Coppock, C\. Ududec, J\. Sekhon, J\. Steinhardt, A\. Kellermann, S\. Schwettmann, M\. Zaharia, I\. Stoica, P\. Liang, and D\. Kang\(2025\)Establishing best practices for building rigorous agentic benchmarks\.External Links:2507\.02825,[Link](https://arxiv.org/abs/2507.02825)Cited by:[1\. Introduction](https://arxiv.org/html/2608.06663#Sx1.p5.1),[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p2.1),[7\. Evaluation & Measurement](https://arxiv.org/html/2608.06663#Sx7.p4.1)\.