Task-Aware Harness Provisioning for LLM Agents in Mission-Critical Infrastructure Operations

# Task-Aware Harness Provisioning for LLM Agents in Mission-Critical Infrastructure Operations
Source: [https://arxiv.org/html/2608.17433](https://arxiv.org/html/2608.17433)
## Task\-Aware Harness Provisioning for LLM Agents in Mission\-Critical Infrastructure OperationsCCS:Computing methodologies Intelligent agentsCCS:Applied computing Physical sciences and engineeringCCS:Computing methodologies Simulation evaluation

Liangtao Lin,Qingang ZhangAffiliation:Nanyang Technological University,Singaporeemail:[qingang\.zhang@ntu\.edu\.sg](mailto:[email protected]),Zhaomeng ZhuAffiliation:Nanyang Technological University,Singaporeemail:[zhaomeng\.zhu@ntu\.edu\.sg](mailto:[email protected]),Tianwei ZhangAffiliation:Nanyang Technological University,Singaporeemail:[tianwei\.zhang@ntu\.edu\.sg](mailto:[email protected])andYonggang WenAffiliation:Nanyang Technological University,Singaporeemail:[ygwen@ntu\.edu\.sg](mailto:[email protected])

© none

###### Abstract\.

LLM agents have been widely adopted to operate mission\-critical infrastructure \(MCI\)\. These agents normally rely on a*harness*that determines what information they can access, which tools they can use, and what actions they can take\. Existing systems often expose the same comprehensive harness to every task, which may not be necessary and cause resource wastes\. In this paper, we focus on the identification of optimal harness configurations, and view it as a resource\-matching problem between what each task requires and what the harness provides\. To measure this match, we classify MCI tasks based on the mathematical representation of the underlying system and rank harness configurations by the amount and type of information they provide\. We then construct task\-to\-harness mappings from two sources: mining research literature and measuring controlled agent execution\. Leveraging the measured mapping, we propose a new harness provisioning algorithm: map\-guided escalation\. It begins with a task\-specific harness and expands to full provision only after a failed self\-check\. We evaluate our method in two representative MCI tasks: in liquid cooling, it improves the agent accuracy from0\.6520\.652under full provision to0\.7150\.715and achieves accuracy comparable to Reflexion with48%48\\%fewer tokens; In power grids, full provision remains accuracy\-optimal, while map\-based provisioning offers lower\-cost alternatives\. These findings show that harness provisioning follows a domain\-dependent accuracy–cost Pareto frontier rather than a universal optimum\.

## 1\.Introduction

Mission\-critical infrastructure \(MCI\), including information technology, power, and water systems, comprises systems and assets whose disruption or destruction could severely affect security, the economy, public health, or public safety\([3](https://arxiv.org/html/2608.17433#bib.bib40)\)\. Operating and maintaining \(O&M\) these systems require continuous monitoring, fault diagnosis, future\-state prediction, maintenance planning, and control\. Recent studies have explored LLM agents for cloud operations and industrial asset management\([19](https://arxiv.org/html/2608.17433#bib.bib23);[28](https://arxiv.org/html/2608.17433#bib.bib24)\)\. To perform these tasks, an agent relies on a*harness*that provides task\-relevant information, tools, and system access\([15](https://arxiv.org/html/2608.17433#bib.bib3);[22](https://arxiv.org/html/2608.17433#bib.bib2)\)\. It is thus critical to configure the optimal harness for each given task\.

We frame this as a resource\-matching problem: a harness determines the information and capabilities available to the agent, while each task imposes its own requirements for reliable execution\. The simplest one\-size\-fits\-all harness strategy, which provides every task with the same configurations, has been questioned\([39](https://arxiv.org/html/2608.17433#bib.bib37)\)\. More advanced solutions are proposed, e\.g\., adapting the harness by retrieving resources relevant to the current task\([29](https://arxiv.org/html/2608.17433#bib.bib6);[24](https://arxiv.org/html/2608.17433#bib.bib7)\), granting access according to predefined rules\([32](https://arxiv.org/html/2608.17433#bib.bib4);[11](https://arxiv.org/html/2608.17433#bib.bib5)\), or allowing the agent to decide which resources it needs\([9](https://arxiv.org/html/2608.17433#bib.bib8);[2](https://arxiv.org/html/2608.17433#bib.bib9)\)\. However, these approaches usually infer what may be useful rather than measure what is actually sufficient\. In MCI O&M, provisioning must balance not only task success but also economic, including token usage, latency, and computation, as well as information security by limiting unnecessary data exposure\. Too little provision may prevent reliable execution, whereas excessive provision increases cost and exposure\. This motivates our central question:can we determine the minimum harness a given task needs for reliable execution?

We hypothesize that each task category has a characteristic harness demand, and seek to identify the mapping between them, which we call the*task\-to\-harness mapping*\. To this end, we first need two comparable representations: one for task demand and one for harness provision\. MCI O&M tasks are commonly described using labels such as detection, diagnosis, and prediction, but these labels lack consistent definitions across domains and studies, and do not systematically cover the full task space\. We therefore derive a new task taxonomy from the mathematical representation of the underlying physical system, yielding categories with explicit and distinguishable boundaries\. On the provision side, we separately define a cumulative harness hierarchy according to the amount and type of system information made available to the agent\. Together, these representations provide two structured spaces for task types and resource provision\.

We propose two complementary strategies to identify the task\-to\-harness mapping\. First, we examine if it is already implicit in prior MCI O&M research\. We collect and analyze more than 1,200 papers published over the past decade and find a clear pattern: tasks concerning future outcomes, latent states, or interventions tend to use more extensive harness resources\. Second, to verify whether this literature\-derived pattern reflects actual agent requirements, we conduct controlled execution experiments in two MCI cases: liquid cooling and power grids\. We build simulation\-backed agent environments with multiple harness levels and a benchmark of 240 verified tasks\. Execution results partly align with the literature\-derived pattern but also reveal domain\-specific variation\. In liquid cooling, several task categories achieve best performance below full provision while using fewer tokens and shorter time, showing that the most comprehensive harness is not always optimal\.

With the task\-to\-harness mapping, an MCI agent can provision resources according to the type and estimated demand of each atomic task\. The simplest solution is to directly retrieve the corresponding harness level from either the literature\-derived or execution\-derived map, requiring no additional routing call\. To account for variation among tasks within the same category, we further propose a*map\-guided escalation*strategy: the agent starts with the execution\-derived provision and retries with the full harness only when the initial execution fails its self\-check\. On held\-out tasks, execution\-map lookup performs comparably to experience\-augmented LLM routing while using14%14\\%and12%12\\%fewer tokens than full provision\. Map\-guided escalation improves accuracy over full provision from0\.6520\.652to0\.7150\.715in liquid cooling\. It also achieves comparable accuracy to Reflexion, an iterative method that uses verbal self\-reflection on task feedback to improve subsequent attempts\([33](https://arxiv.org/html/2608.17433#bib.bib16)\), while using48%48\\%fewer tokens\. In the power\-grid domain, full provision remains accuracy\-optimal, while execution\-map lookup provides a lower\-cost operating point\. Overall, these results reveal a domain\-dependent Pareto frontier between execution accuracy and resource cost rather than a single provisioning policy that is optimal across systems\.

Our contributions are summarized as follows:

- •We formulate harness provisioning for MCI agents as a resource\-matching problem, balancing task performance, execution cost, and unnecessary information exposure\.
- •We introduce a novel MCI agent task taxonomy derived from system equations and an ordered harness hierarchy with increasing system access, providing comparable representations of task demand and harness provision\.
- •We construct two task\-to\-harness maps through data\-driven analysis of more than 1,200 MCI papers and controlled execution on a new benchmark of 240 verified tasks in two physical cases\.
- •We propose map\-guided escalation, which uses the measured mapping to initialize harness provision and expands access only when needed, improving the accuracy–cost trade\-off over fixed and dynamically routed provisioning\.

![Refer to caption](https://arxiv.org/html/2608.17433v1/figure1.png)Figure 1\.Overview of the Proposed Framework, which defines the task and harness spaces, estimates literature\- and execution\-derived task\-to\-harness maps, and uses the resulting map for harness selection with optional self\-check\-triggered escalation\.
## 2\.Related Work

### 2\.1\.LLM Agents and Benchmarks for MCI

Operational intelligence in mission\-critical infrastructure has traditionally relied on task\-specific pipelines\. Representative KDD systems include EGADS for scalable anomaly detection\([13](https://arxiv.org/html/2608.17433#bib.bib20)\), SR\-CNN and OmniAnomaly for time\-series monitoring\([31](https://arxiv.org/html/2608.17433#bib.bib21);[34](https://arxiv.org/html/2608.17433#bib.bib31)\), and CIRCA for causal root\-cause analysis\([16](https://arxiv.org/html/2608.17433#bib.bib22)\)\. These methods provide strong solutions for individual detection or diagnosis tasks, but do not support a general agent across the O&M lifecycle\.

Recent work moves toward interactive agents and executable benchmarks\. AIOpsLab deploys fault\-injected cloud environments and evaluates agents across the incident lifecycle\([19](https://arxiv.org/html/2608.17433#bib.bib23)\), while AssetOpsBench provides tools and scenarios for industrial asset operation and maintenance\([28](https://arxiv.org/html/2608.17433#bib.bib24)\)\. In related infrastructure control, LLMLight evaluates LLM agents for traffic\-signal decision making\([12](https://arxiv.org/html/2608.17433#bib.bib32)\)\. Existing surveys emphasize that realistic agent evaluation must cover behavior, reliability, safety, environments, and tooling\([23](https://arxiv.org/html/2608.17433#bib.bib33)\)\. These benchmarks evaluate agents within predefined environments and interfaces\. Our benchmark instead varies harness provision along an ordered ladder, holds the executor fixed, and estimates the sufficient provision for each task class\.

### 2\.2\.Improving Agent Performance

Most prior work improves the execution side of an agent by changing how it reasons, learns, or reacts within a given environment\. ReAct interleaves reasoning with actions\([37](https://arxiv.org/html/2608.17433#bib.bib15)\), Reflexion introduces feedback\-driven retries\([33](https://arxiv.org/html/2608.17433#bib.bib16)\), and ExpeL retrieves experience from prior trajectories\([41](https://arxiv.org/html/2608.17433#bib.bib17)\)\. Training\-based methods further strengthen planning and tool use: AgentGen generates environments and tasks for planning\-oriented instruction tuning\([8](https://arxiv.org/html/2608.17433#bib.bib25)\), while Tool\-MVR learns tool invocation and error correction from verified trajectories\([20](https://arxiv.org/html/2608.17433#bib.bib26)\)\. These methods primarily improve the executor or its execution process under an available harness\.

A complementary line of work improves the supply side by adapting what is exposed to the executor\. ToolLLM and AnyTool retrieve task\-relevant APIs from large tool collections\([29](https://arxiv.org/html/2608.17433#bib.bib6);[4](https://arxiv.org/html/2608.17433#bib.bib11)\), while Instruction\-Tool Retrieval \(ITR\) dynamically retrieves relevant instruction fragments and exposes a reduced tool subset at each step\([6](https://arxiv.org/html/2608.17433#bib.bib34)\)\. Chameleon plans compositions of external modules\([18](https://arxiv.org/html/2608.17433#bib.bib12)\), and Sufficient Context predicts whether retrieved textual evidence is adequate for answering a query\([10](https://arxiv.org/html/2608.17433#bib.bib10)\)\. Model\-routing methods such as RouteLLM and AutoMix similarly allocate computational capacity according to predicted need, although they switch executors rather than vary the harness of a fixed executor\([25](https://arxiv.org/html/2608.17433#bib.bib13);[1](https://arxiv.org/html/2608.17433#bib.bib14)\)\. These approaches select resources through relevance, learned routing, or model\-reported need\. Our work instead measures the lowest harness provision that preserves execution performance\.

### 2\.3\.Harness Design, Evaluation, and Governance

Recent work increasingly treats the agent harness as a first\-class system layer spanning execution control, tool access, context, state, verification, and recovery\([15](https://arxiv.org/html/2608.17433#bib.bib3);[22](https://arxiv.org/html/2608.17433#bib.bib2)\)\. Natural\-Language Agent Harnesses externalizes control logic into portable specifications\([27](https://arxiv.org/html/2608.17433#bib.bib27)\)\. Meta\-Harness searches over harness implementations\([14](https://arxiv.org/html/2608.17433#bib.bib28)\)\. Harness\-Bench measures configuration\-level harness effects across models and workflows\([38](https://arxiv.org/html/2608.17433#bib.bib29)\)\. These studies establish that agent performance depends on the model–harness configuration, but focus on harness representation, optimization, or cross\-configuration evaluation\.

A parallel body of work governs how harness resources may be used\. Progent enforces programmable tool policies\([32](https://arxiv.org/html/2608.17433#bib.bib4)\), Prompt Flow Integrity constrains information flows\([11](https://arxiv.org/html/2608.17433#bib.bib5)\), and AgentSandbox and MiniScope provide sandboxing and permission analysis\([40](https://arxiv.org/html/2608.17433#bib.bib18);[42](https://arxiv.org/html/2608.17433#bib.bib19)\)\. ToolPrivBench studies unnecessary privilege selection\([36](https://arxiv.org/html/2608.17433#bib.bib1)\), while HarnessAudit evaluates boundary compliance over complete execution trajectories\([17](https://arxiv.org/html/2608.17433#bib.bib30)\)\.Collectively, these studies treat the harness as an object to design, optimize, evaluate, or govern\. Our work addresses a complementary question: for a fixed executor and recurring task class, what is the sufficient harness provision that preserves execution performance?

## 3\.Overview

Figure[1](https://arxiv.org/html/2608.17433#S1.F1)presents the overall workflow of our methodology, which consists of three stages: defining the task and harness spaces, estimating the relationship between them, and using the resulting mapping to select harnesses for new tasks\.

First, to compare what different tasks require with what different harnesses provide, we formally define the task space from the underlying physical system, and organize harnesses into levels with increasing information and capabilities \(Section[4](https://arxiv.org/html/2608.17433#S4)\)\.

Second, we establish the task\-to\-harness mapping from two perspectives \(Section[5](https://arxiv.org/html/2608.17433#S5)\)\. We first mine existing MCI O&M studies to examine which harness levels prior work uses for different task classes\. We then construct a controlled benchmark in two MCI environments and execute the same agent under different harness levels to empirically estimate the mapping\.

Third, we investigate how the resulting map can be used to select harnesses for new tasks \(Section[6](https://arxiv.org/html/2608.17433#S6)\)\. Given a MCI agent task, we first assign it to a task class and retrieve the corresponding harness level from the map\. Direct lookup uses this level as the final provision, whereas map\-guided escalation uses it as the initial provision and retries with the full harness if the first execution fails its self\-check\. This allows the class\-level map to guide efficient provisioning while accounting for variation among individual tasks\.

## 4\.Task–Harness Spaces

Determining which harness a task requires is fundamentally a measurement problem\. Before this relationship can be measured, task demand and harness provision must be expressed using stable and comparable representations\. We therefore construct two rulers\. The first organizes MCI tasks according to what they query or affect in the underlying physical system\. The second organizes harness configurations according to the information, tools, and operational access provided to the agent\. Together, these rulers turn harness provisioning into a measurable mapping from task class to minimum sufficient harness\.

### 4\.1\.Task Space

Conventional task labels such as detection, diagnosis, prediction, and planning are unsuitable as primary task categories because their meanings vary across domains, applications, and implementations\. We therefore derive the task space directly from a common representation of partially observed physical systems\.

Letyty\_\{t\}denote observable signals,xtx\_\{t\}denote latent physical states, andmtm\_\{t\}denote the system structure and governing mechanisms, including topology, dynamics, constraints, and control logic\. We abstract the system as

xt\+1=Fmt​\(xt,at,dt\)\+ξt,yt=Gmt​\(xt\)\+ϵt,x\_\{t\+1\}=F\_\{m\_\{t\}\}\(x\_\{t\},a\_\{t\},d\_\{t\}\)\+\\xi\_\{t\},\\qquad y\_\{t\}=G\_\{m\_\{t\}\}\(x\_\{t\}\)\+\\epsilon\_\{t\},whereata\_\{t\}is an intervention,dtd\_\{t\}is an external disturbance, andξt\\xi\_\{t\}andϵt\\epsilon\_\{t\}denote process and observation uncertainty\. This representation separates what is directly observed, what remains latent, and what governs system behavior\. These three elements provide the possible targets of an MCI task\.

Then we represent each task as

T=⟨ω,τ,e⟩,ω∈\{Inform,Act\},τ∈\{t,t\+Δ\},e∈\{y,x,m\}\.T=\\langle\\omega,\\tau,e\\rangle,\\quad\\omega\\in\\\{\\mathrm\{Inform\},\\mathrm\{Act\}\\\},\\tau\\in\\\{t,t\+\\Delta\\\},e\\in\\\{y,x,m\\\}\.The target elementeespecifies whether the task concerns observable signals, latent states, or system mechanisms\. The target timeτ\\taudistinguishes the current system from a future outcome\. The output modeω\\omegaspecifies whether the agent should report information about the target or produce an intervention intended to affect it\. Their Cartesian product yields2×2×3=122\\times 2\\times 3=12canonical task classes\.

Given the evidenceBtB\_\{t\}available at task time, the required output takes one of two forms:

oT=\{r^T=ρT​\(p⁡\(eτ∣Bt\)\),ω=Inform,a⋆=arg⁡mina∈𝒜⁡𝔼⁡\[JT​\(eτd​o​\(a\)\)∣Bt\],ω=Act\.o\_\{T\}=\\begin\{cases\}\\displaystyle\\hat\{r\}\_\{T\}=\\rho\_\{T\}\\\!\\left\(p\(e\_\{\\tau\}\\mid B\_\{t\}\)\\right\),&\\omega=\\mathrm\{Inform\},\\\\\[10\.0pt\] \\displaystyle a^\{\\star\}=\\arg\\min\_\{a\\in\\mathcal\{A\}\}\\mathbb\{E\}\\\!\\left\[J\_\{T\}\\\!\\left\(e\_\{\\tau\}^\{do\(a\)\}\\right\)\\mid B\_\{t\}\\right\],&\\omega=\\mathrm\{Act\}\.\\end\{cases\}Here,p⁡\(eτ∣Bt\)p\(e\_\{\\tau\}\\mid B\_\{t\}\)represents the agent’s inferred belief about the target, andρT\\rho\_\{T\}converts this belief into the task\-specific reportr^T\\hat\{r\}\_\{T\}, such as a value, label, event, or explanation\. For an Act task,eτd​o​\(a\)e\_\{\\tau\}^\{do\(a\)\}denotes the target outcome under interventionaa, andJTJ\_\{T\}evaluates the desirability of that outcome\. An Inform task therefore reports knowledge abouteeat timeτ\\tau, whereas an Act task selects an intervention according to its expected effect oneeat that time\.

Conventional task types can be represented by specific coordinates in this space according to their target, time horizon, and output mode\. For example, detection corresponds to⟨Inform,t,y⟩\\langle\\mathrm\{Inform\},t,y\\rangle, diagnosis to⟨Inform,t,m⟩\\langle\\mathrm\{Inform\},t,m\\rangle, and prediction to⟨Inform,t\+Δ,y⟩\\langle\\mathrm\{Inform\},t\+\\Delta,y\\rangle\. In the following, we use the three\-dimensional representation⟨ω,τ,e⟩\\langle\\omega,\\tau,e\\rangleto denote task types\.

### 4\.2\.Harness Space

A task representation specifies what must be accomplished, but not what information, models, tools, and interfaces are exposed to the executor\. Motivated by prior work that organizes system information by temporal reach and structural scope\([5](https://arxiv.org/html/2608.17433#bib.bib38);[30](https://arxiv.org/html/2608.17433#bib.bib39)\), we define five cumulative levels of harness provision:

ℋK​1⊂ℋK​2⊂ℋK​3⊂ℋK​4⊂ℋK​5\.\\mathcal\{H\}\_\{K1\}\\subset\\mathcal\{H\}\_\{K2\}\\subset\\mathcal\{H\}\_\{K3\}\\subset\\mathcal\{H\}\_\{K4\}\\subset\\mathcal\{H\}\_\{K5\}\.
- •K​1K1: Model\-only reasoning\.The agent receives only the task prompt and relies on its parametric knowledge\.
- •K​2K2: Static knowledge\.The harness additionally provides fixed resources such as manuals, SOPs, specifications, design documents, and rule bases\.
- •K​3K3: Temporal observations\.The harness additionally provides historical or real\-time telemetry, logs, alarms, and time\-series measurements\.
- •K​4K4: Structure and physics\.The harness additionally provides topology, component relations, governing equations, physical constraints, and control logic\.
- •K​5K5: Forward simulation\.The harness additionally provides executable simulation, digital\-twin rollouts, counterfactual evaluation, and optimization over hypothetical future trajectories\.

The ordering reflects increasing access to the physical system rather than increasing intrinsic intelligence of the executor\. It also does not impose a fixed mapping from task coordinates to harness levels: a future\-facing task may be solved from temporal observations, while a current\-state task may require structural models or simulation\. Although provision increases fromK​1K1toK​5K5, performance need not improve monotonically because additional context and tools may increase cost, introduce irrelevant evidence, or create unnecessary action opportunities\. Thus,K​5K5denotes maximal provision rather than an assumed performance optimum\.

## 5\.Estimating the Task\-to\-Harness Map

The task space characterizes what different MCI tasks demand, while the harness space characterizes what information and capabilities are provided to the agent\. Our goal is to identify the relationship between these two spaces by assigning a harness level to each task class\. We first define a common mapping rule and then estimate the map from two sources\. The literature\-derived map summarizes which harness levels prior MCI O&M studies use for different task classes, whereas the execution\-derived map identifies the lowest harness level that achieves near\-best performance when the same agent is evaluated under different harness levels\.

### 5\.1\.Mapping Definition

LetAs​\(T,K\)A\_\{s\}\(T,K\)denote the support for assigning harnessℋK\\mathcal\{H\}\_\{K\}to task classTTunder evidence sources∈\{lit,exec\}s\\in\\\{\\mathrm\{lit\},\\mathrm\{exec\}\\\}\. We define

𝒢s​\(T\)=min⁡\{K∈𝒦:As​\(T,K\)≥maxK′∈𝒦⁡As​\(T,K′\)−ϵs\},\\mathcal\{G\}\_\{s\}\(T\)=\\min\\left\\\{K\\in\\mathcal\{K\}:A\_\{s\}\(T,K\)\\geq\\max\_\{K^\{\\prime\}\\in\\mathcal\{K\}\}A\_\{s\}\(T,K^\{\\prime\}\)\-\\epsilon\_\{s\}\\right\\\},whereϵs≥0\\epsilon\_\{s\}\\geq 0is a source\-specific tolerance\. The mapping selects the lowest harness level whose support is withinϵs\\epsilon\_\{s\}of the maximum for that task class, avoiding unnecessary provision when several levels receive similar support\.

For the literature\-derived map, we setAlit​\(T,K\)=Plit​\(K∣T\)A\_\{\\mathrm\{lit\}\}\(T,K\)=P\_\{\\mathrm\{lit\}\}\(K\\mid T\), wherePlit​\(K∣T\)P\_\{\\mathrm\{lit\}\}\(K\\mid T\)is the observed frequency of harness levelKKamong prior studies assigned to task classTT\. The resulting map𝒢lit\\mathcal\{G\}\_\{\\mathrm\{lit\}\}summarizes how prior work has provisioned different task classes\.

For the execution\-derived map, we setAexec​\(T,K\)=μexec​\(T,K\)A\_\{\\mathrm\{exec\}\}\(T,K\)=\\mu\_\{\\mathrm\{exec\}\}\(T,K\), whereμexec​\(T,K\)\\mu\_\{\\mathrm\{exec\}\}\(T,K\)is the mean execution score obtained under harness levelKKfor task classTT\. The resulting map𝒢exec\\mathcal\{G\}\_\{\\mathrm\{exec\}\}selects the lowest harness level whose measured performance is withinϵexec\\epsilon\_\{\\mathrm\{exec\}\}of the best observed performance\.

### 5\.2\.Literature\-Derived Mapping

We first examine which harness levels existing MCI O&M studies use for different task classes\. We collect approximately 2,000 candidate papers fromarXiv,Semantic Scholar, andOpenAlex, and retain more than 1,200 studies within our scope\. For each paperpp, three LLM annotators independently extract its primary task classTp∈𝒯T\_\{p\}\\in\\mathcal\{T\}and the highest harness levelKp∈𝒦K\_\{p\}\\in\\mathcal\{K\}materially used by the proposed method\. The labels are consolidated by majority vote, with unresolved disagreements manually adjudicated\.111Appendix[A](https://arxiv.org/html/2608.17433#A1)details the collection, screening, and annotation procedures\.

We then estimate

Plit​\(K∣T\)=Nlit​\(T,K\)∑K′∈𝒦Nlit​\(T,K′\),P\_\{\\mathrm\{lit\}\}\(K\\mid T\)=\\frac\{N\_\{\\mathrm\{lit\}\}\(T,K\)\}\{\\sum\_\{K^\{\\prime\}\\in\\mathcal\{K\}\}N\_\{\\mathrm\{lit\}\}\(T,K^\{\\prime\}\)\},whereNlit​\(T,K\)N\_\{\\mathrm\{lit\}\}\(T,K\)is the paper count associated with task classTTand harness levelKK\. Applying the common mapping rule produces𝒢lit\\mathcal\{G\}\_\{\\mathrm\{lit\}\}\.

The above map captures established provisioning practice rather than verified sufficiency\. A frequently\-used harness level may still be unnecessary, insufficient, or over\-provisioned for a particular agent or environment\. Nevertheless, it provides a literature\-scale view of how harness provision varies across task classes\.

Figure 2\.Literature\-derived Harness Distributions across the 12 Task Classes\.Based on 1,220 MCI O&M papers, each panel shows the number of papers assigned to harness levelsK1K\_\{1\}–K5K\_\{5\}for one task class; the outlined bar marks the modal level while the triangle and background marks the mean\. The distributions generally shift toward higher harness levels for Act, Future, and latent\-state tasks, while most classes peak at eitherK3K\_\{3\}orK5K\_\{5\}\.
### 5\.3\.Execution\-Derived Mapping

We next measure the relationship directly by evaluating the same agent under different harness levels222Appendix[C](https://arxiv.org/html/2608.17433#A3)details the domain\-specific harness implementation\.\. We construct two simulation\-backed environments: a thermal–hydraulic digital twin of a liquid\-cooled data hall and a modified IEEE\-14 power\-grid environment built with Grid2Op and pandapower\([21](https://arxiv.org/html/2608.17433#bib.bib42);[35](https://arxiv.org/html/2608.17433#bib.bib41)\)\. Each environment generates replayable system trajectories from which we instantiate ten tasks for each of the 12 task classes, yielding 240 tasks in total\. The tasks cover reporting current observations, inferring latent states or system mechanisms, predicting future outcomes, and selecting interventions\. Each task is paired with a deterministic simulator\-backed oracle and undergoes both programmatic and human verification\.333Appendix[B](https://arxiv.org/html/2608.17433#A2)details the benchmark construction and verification\.

Every task is executed under all five harness levels using the same agent, reasoning protocol, and evaluation procedure; only the available harness provision varies\. Letμexec​\(T,K\)\\mu\_\{\\mathrm\{exec\}\}\(T,K\)denote the mean score of harness levelKKover tasks in classTT\. Applying the common mapping rule gives

𝒢exec​\(T\)=min⁡\{K∈𝒦:μexec​\(T,K\)≥maxK′∈𝒦⁡μexec​\(T,K′\)−ϵexec\}\.\\mathcal\{G\}\_\{\\mathrm\{exec\}\}\(T\)=\\min\\left\\\{K\\in\\mathcal\{K\}:\\mu\_\{\\mathrm\{exec\}\}\(T,K\)\\geq\\max\_\{K^\{\\prime\}\\in\\mathcal\{K\}\}\\mu\_\{\\mathrm\{exec\}\}\(T,K^\{\\prime\}\)\-\\epsilon\_\{\\mathrm\{exec\}\}\\right\\\}\.
The above execution\-derived map identifies the lowest harness level that achieves near\-best performance when the same agent is evaluated across harness levels\.

![Refer to caption](https://arxiv.org/html/2608.17433v1/figure3.png)Figure 3\.Task\-to\-Harness maps\.Cells report the mean construction\-split accuracy for each task class and harness level \(5 tasks per class, 3 runs\)\. Amber boxes mark the lowest level withinϵexec=0\.05\\epsilon\_\{\\mathrm\{exec\}\}=0\.05of the best execution score, while the black staircase marks the literature\-derived level\. Columns are ordered by the literature mean harness level\.

## 6\.Map\-Guided Harness Selection

Given an atomic taskqq, the system first assigns it to a task classT⁡\(q\)T\(q\)and uses the task\-to\-harness map to select an initial harness\. The executor then completes the task using only the information and capabilities exposed by the selected harness\. Because tasks within the same class may still differ in difficulty and required information, we allow the harness to expand toK5K\_\{5\}when the initial execution fails its self\-check\.

Initial harness selection\.Since the task\-to\-harness map captures recurring provisioning patterns for each task class, we use the mapped level as a class\-level prior for selecting the initial harness\. Given a taskqqand a map𝒢s\\mathcal\{G\}\_\{s\}, wheres∈\{lit,exec\}s\\in\\\{\\mathrm\{lit\},\\mathrm\{exec\}\\\}denotes the map source, the initial harness level is

K0​\(q,s\)=𝒢s​\(T⁡\(q\)\)\.K\_\{0\}\(q;s\)=\\mathcal\{G\}\_\{s\}\(T\(q\)\)\.Once the task classT⁡\(q\)T\(q\)is available, this step requires only a deterministic table lookup and introduces no additional routing call\. Letπ\\pidenote the fixed executor\. Runningπ\\piunderℋK0​\(q,s\)\\mathcal\{H\}\_\{K\_\{0\}\(q;s\)\}produces

\(o0,c0,r0\)=π⁡\(q,ℋK0​\(q,s\)\),\(o\_\{0\},c\_\{0\},r\_\{0\}\)=\\pi\\\!\\left\(q,\\mathcal\{H\}\_\{K\_\{0\}\(q;s\)\}\\right\),whereo0o\_\{0\}is the initial output,c0∈\{pass,fail\}c\_\{0\}\\in\\\{\\mathrm\{pass\},\\mathrm\{fail\}\\\}is its self\-check result, andr0r\_\{0\}contains findings that can be reused in a subsequent attempt\.

Conditional escalation\.The mapped level is a class\-level prior and may not provide sufficient information or capabilities for every task instance\. When the initial execution fails its self\-check, we therefore expand the harness toK5K\_\{5\}, exposing the executor to the full set of information, tools, and operational permissions\. The initial result is accepted if it passes the self\-check or if the selected harness is alreadyK5K\_\{5\}; otherwise, the executor retries once underK5K\_\{5\}while retaining the useful findingsr0r\_\{0\}from the initial attempt\. The final harness and output are

\(Kfinal​\(q\),o⁡\(q\)\)=\{\(K0​\(q,s\),o0\),c0=pass​or​K0​\(q,s\)=K5,\(K5,π⁡\(q,ℋK5,r0\)\),c0=fail​and​K0​\(q,s\)<K5\.\\bigl\(K\_\{\\mathrm\{final\}\}\(q\),o\(q\)\\bigr\)=\\begin\{cases\}\\bigl\(K\_\{0\}\(q;s\),o\_\{0\}\\bigr\),&c\_\{0\}=\\mathrm\{pass\}\\ \\text\{or\}\\ K\_\{0\}\(q;s\)=K\_\{5\},\\\\\[3\.0pt\] \\bigl\(K\_\{5\},\\pi\(q,\\mathcal\{H\}\_\{K\_\{5\}\};r\_\{0\}\)\\bigr\),&c\_\{0\}=\\mathrm\{fail\}\\ \\text\{and\}\\ K\_\{0\}\(q;s\)<K\_\{5\}\.\\end\{cases\}
Method variants\.This formulation produces three evaluated variants\.Lit\-Lookupuses𝒢lit\\mathcal\{G\}\_\{\\mathrm\{lit\}\}and disables escalation, whileExec\-Lookupuses𝒢exec\\mathcal\{G\}\_\{\\mathrm\{exec\}\}and likewise accepts the mapped harness as final\.Map\-ESCuses𝒢exec\\mathcal\{G\}\_\{\\mathrm\{exec\}\}for initial selection and applies the self\-check\-triggered escalation rule above\. The two lookup variants isolate the value of the maps themselves, whileMap\-ESCevaluates whether failures caused by instance\-level variation can be recovered through conditional expansion\.

We escalate directly toK5K\_\{5\}rather than testing each intermediate level to avoid repeated execution overhead; Appendix[H\.2](https://arxiv.org/html/2608.17433#A8.SS2)shows that intermediate retries rarely terminate before reachingK5K\_\{5\}in the evaluated domains\. The executor and reasoning protocol remain fixed across both attempts, and the harness enforces the capability boundary at each level\. Before escalation, the executor therefore cannot access information, tools, or permissions beyond its initially selected harness\.

## 7\.Experiments

Having defined the task\-to\-harness maps and their deployment policies, we now evaluate whether the proposed structure is empirically supported and operationally useful\. Our experiments address four research questions\.

- •RQ1: Do task classes exhibit distinct harness requirements? Does richer provision consistently improve performance?
- •RQ2: How reliably can the task\-to\-harness map be estimated from literature and execution? How do the two evidence sources compare with each other?
- •RQ3: How should the map be operationalized at deployment time: through direct lookup or map\-guided escalation?
- •RQ4: Whether do the resulting task\-to\-harness relationships generalize across domains and executor models? What is the impact of the granularity, robustness, and deployment implications of the learned map?

Table 1\.Harness Provisioning on the Test Split\.All policies use the same frozen executor and differ only in how the harness is selected\. Latency and tokens are normalized to Full\-K5\. Bold and underline denote the best and second\-best deployable results in each column\.### 7\.1\.Experimental Setup

Environments and tasks\.We evaluate two simulation\-backed MCI domains: a proprietary liquid\-cooling digital twin \(Liquid\) and a power\-grid environment built with Grid2Op and pandapower \(Grid\)\. Each domain contains 120 tasks, with ten instances for each of the 12 classesT=⟨ω,τ,e⟩T=\\langle\\omega,\\tau,e\\rangle\. Within each class, five tasks are used for map construction and five form a disjoint test split\. All 240 tasks undergo programmatic and human verification\. We perform the fullK1K\_\{1\}–K5K\_\{5\}sweep on all tasks, but construct the execution\-derived map only from the construction split and evaluate provisioning policies on the test split\.

Execution protocol\.We instantiate the five cumulative harness levels defined in Section[4](https://arxiv.org/html/2608.17433#S4)\. Unless otherwise stated, all conditions use the same frozen GPT\-5\.4\([26](https://arxiv.org/html/2608.17433#bib.bib35)\)executor with a ReAct loop, at most 20 tool iterations, and a 300\-second timeout; only harness provisioning varies\. Both maps use a tolerance ofϵlit=ϵexec=0\.05\\epsilon\_\{\\mathrm\{lit\}\}=\\epsilon\_\{\\mathrm\{exec\}\}=0\.05\.444Appendix[E](https://arxiv.org/html/2608.17433#A5)reports tolerance sensitivity, selected\-level stability, statistical tests, and oracle definitions\.

Compared policies\.We compare maximal provisioning \(Full\-K5K\_\{5\}\), relevance\-based retrieval \(ITR\([6](https://arxiv.org/html/2608.17433#bib.bib34)\)\), LLM routing with and without prior experience \(LLM\-Route, LLM\+Exp\), a learned cascade \(AutoMix\([1](https://arxiv.org/html/2608.17433#bib.bib14)\)\), and progressive escalation fromK1K\_\{1\}\(Blind\-ESC\)\. Our policies include direct lookup from the literature and execution maps \(Lit\-Lookup, Exec\-Lookup\) and two\-stage map\-guided escalation from the execution\-derived mapped level \(Map\-ESC\)\. We separately compare Map\-ESC with Reflexion\([33](https://arxiv.org/html/2608.17433#bib.bib16)\)and ExpeL\([41](https://arxiv.org/html/2608.17433#bib.bib17)\), which adapt execution under a fixedK5K\_\{5\}harness\.555Appendix[F](https://arxiv.org/html/2608.17433#A6)details the implementation and adaptation of all compared policies\.

Metrics\.Task performance is scored in\[0,1\]\[0,1\]using a Gemini\-3\.1\-Pro\([7](https://arxiv.org/html/2608.17433#bib.bib36)\)judge combined with rule\-based checks\. Each condition is run three times, and results are reported as mean±\\pmpopulation standard deviation\. Token counts include routing calls and all escalation attempts; latency and tokens are normalized to Full\-K5K\_\{5\}\. We report two non\-deployable references: theClass Oracleselects the highest\-scoring harness for each task class, while theTask Oracleselects the highest\-scoring harness separately for each task, both using mean scores across the three runs\.666Appendix[D](https://arxiv.org/html/2608.17433#A4)details the executor settings, scoring rules, judge configuration, cost accounting, and repetition protocol\.

### 7\.2\.RQ1: Do Task Classes Require Different Harnesses?

Figure[2](https://arxiv.org/html/2608.17433#S5.F2)shows that harness provision in prior MCI O&M studies varies systematically across task classes\. Among 1,200\+ papers, current Inform tasks targeting observable signals or system mechanisms are concentrated atK3K\_\{3\}, whereas Future, Act, and latent\-state tasks shift more strongly towardK5K\_\{5\}\. This indicates two recurring provisioning patterns in prior work: observation\-driven provision centred onK3K\_\{3\}and simulation\-backed provision centred onK5K\_\{5\}\.

Figure[3](https://arxiv.org/html/2608.17433#S5.F3)provides direct execution evidence by evaluating the same executor underK1K\_\{1\}–K5K\_\{5\}\.777Appendix[G](https://arxiv.org/html/2608.17433#A7)details the aggregate and per\-classK1K\_\{1\}–K5K\_\{5\}sweeps, split\-specific analyses, and absolute execution costs\.The lowest level achieving performance withinϵ=0\.05\\epsilon=0\.05of the best class score varies fromK2K\_\{2\}toK5K\_\{5\}inLiquidand fromK3K\_\{3\}toK5K\_\{5\}inGrid\. Performance is also non\-monotonic: five classes inLiquidand four inGridattain their highest score belowK5K\_\{5\}\. Thus, harness requirements vary across both task classes and domains, and maximal provision is neither uniformly necessary nor uniformly optimal\.

### 7\.3\.RQ2: How Should the Task\-to\-Harness Map Be Estimated?

Having established that harness requirements vary across task classes, we next compare the two sources used to estimate the map\. As shown in Table[1](https://arxiv.org/html/2608.17433#S7.T1), Lit\-Lookup selects the level most frequently observed in prior studies, whereas Exec\-Lookup selects the lowest level whose construction\-split performance is withinϵexec=0\.05\\epsilon\_\{\\mathrm\{exec\}\}=0\.05of the best level for each task class\. The two maps perform similarly onLiquid\(0\.6580\.658versus0\.6700\.670\), but differ substantially onGrid, where Exec\-Lookup achieves0\.7620\.762compared with0\.6640\.664\. Literature evidence therefore provides a useful prior, while execution evidence better adapts the map to the evaluated domain and executor\.

Exec\-Lookup also provides a stronger accuracy–efficiency balance than per\-task routing\. Directly asking an LLM to select the harness \(LLM\-Route\) yields lower accuracy in both domains \(0\.6260\.626and0\.7440\.744\)\. Providing the router with a playbook distilled from construction\-split trajectories \(LLM\+Exp\) improves its decisions \(0\.6550\.655and0\.7660\.766\), but does not consistently outperform Exec\-Lookup and incurs an additional routing call\. Retrieving resources according to their semantic relevance to the task \(ITR\) achieves comparable accuracy \(0\.6700\.670and0\.7800\.780\) but remains close to Full\-K5K\_\{5\}in token usage \(0\.99×0\.99\\timesand1\.02×1\.02\\times\), compared with0\.86×0\.86\\timesand0\.88×0\.88\\timesfor Exec\-Lookup\. Semantic relevance can therefore identify potentially useful resources, but does not directly determine a cost\-efficient harness\. Overall, measured class\-level execution evidence provides the most consistent basis for initial harness selection\.

Figure 4\.Accuracy–Cost Trade\-offs of Harness Provisioning Policies on the Test Split\. Latency and tokens are normalized to Full\-K5K\_\{5\}; dashed lines show the empirical Pareto frontiers\.
### 7\.4\.RQ3: How Should the Map Guide Harness Selection?

The task\-to\-harness map supports two deployment modes: using the mapped harness directly or using it as the initial provision before conditional escalation\. As shown in Table[1](https://arxiv.org/html/2608.17433#S7.T1)and Figure[4](https://arxiv.org/html/2608.17433#S7.F4), Exec\-Lookup achieves accuracies of0\.6700\.670onLiquidand0\.7620\.762onGrid, while reducing token usage to0\.86×0\.86\\timesand0\.88×0\.88\\timesrelative to Full\-K5K\_\{5\}\. Map\-ESC retries underK5K\_\{5\}when the initial execution fails its self\-check, increasing accuracy to0\.7150\.715and0\.7820\.782, with token costs of1\.15×1\.15\\timesand1\.03×1\.03\\times, respectively\. Direct lookup therefore favors efficiency, while conditional escalation improves accuracy at additional cost\.

Blind\-ESC and AutoMix represent two alternative escalation strategies\. Blind\-ESC starts fromK1K\_\{1\}and progressively expands the harness based on self\-checks, whereas AutoMix uses a learned cascade to determine whether further provision is needed\. Map\-ESC instead selects the initial harness from the execution\-derived map and allows at most one fallback toK5K\_\{5\}\. OnLiquid, Map\-ESC outperforms both Blind\-ESC \(0\.7150\.715versus0\.6640\.664\) and AutoMix \(0\.7150\.715versus0\.6390\.639\)\. OnGrid, it outperforms Blind\-ESC \(0\.7820\.782versus0\.7420\.742\) and nearly matches AutoMix \(0\.7820\.782versus0\.7850\.785\), while using fewer tokens than both \(1\.03×1\.03\\timesversus1\.75×1\.75\\timesand2\.61×2\.61\\times\)\. Map\-guided two\-stage escalation therefore provides the most consistent accuracy–efficiency trade\-off across the two domains\.

Figure[4](https://arxiv.org/html/2608.17433#S7.F4)further shows that choosing an appropriate harness configuration depends on both the domain and the deployment objective\. OnLiquid, Exec\-Lookup provides a low\-cost configuration, while Map\-ESC achieves the highest accuracy among the harness\-selection methods\. OnGrid, Full\-K5K\_\{5\}remains the most accurate configuration, whereas Exec\-Lookup offers a lower\-cost Pareto\-efficient alternative\. The map therefore helps determine when additional provision improves performance and when it only increases cost\.

Table[2](https://arxiv.org/html/2608.17433#S7.T2)compares Map\-ESC with execution\-side methods operating under a fixedK5K\_\{5\}harness\. Map\-ESC matches Reflexion in accuracy \(0\.7150\.715versus0\.7110\.711\) while using48%48\\%fewer tokens\. ExpeL improves accuracy by0\.0160\.016but requires44%44\\%more tokens\. Harness provisioning and execution\-side adaptation are therefore complementary, with measured provisioning offering the stronger efficiency trade\-off in this setting\.888Appendix[H](https://arxiv.org/html/2608.17433#A8)details routing\-granularity results, escalation diagnostics, experience ablations, and qualitative error analysis\.

Table 2\.Provisioning Versus Execution\-side Adaptation on theLiquidTest Split\. Tokens are normalized to Map\-ESC\.
### 7\.5\.RQ4: Does the Map Generalize?

The two domains exhibit the same qualitative need for task\-aware provisioning but different deployment outcomes\. As shown in Table[1](https://arxiv.org/html/2608.17433#S7.T1)and Figure[4](https://arxiv.org/html/2608.17433#S7.F4), task\-aware harness selection improves both accuracy and cost inLiquid, whereas Full\-K5K\_\{5\}remains accuracy\-optimal inGridand Exec\-Lookup provides a lower\-cost Pareto\-efficient alternative\. Task\-dependent harness demand therefore persists across domains, although the benefit of reducing provision is domain\-specific\.

We further transfer the execution\-derived map estimated with GPT\-5\.4 to Qwen3\.5\-27B without recalibration\. As shown in Table[3](https://arxiv.org/html/2608.17433#S7.T3), the transferred map improves Qwen over its Full\-K5K\_\{5\}baseline onLiquid\(0\.7260\.726versus0\.7060\.706\) and remains close onGrid\(0\.7970\.797versus0\.8080\.808\)\. In contrast, a map estimated from a single Qwen run performs substantially worse in both domains\. The task\-to\-harness relationship therefore transfers across executors, while reliable map estimation benefits from stronger and repeated measurements\.

Table 3\.Cross\-executor Transfer of the Execution\-derived Map\. All conditions use Qwen3\.5\-27B as the executor\.

## 8\.Conclusion

We formulate harness provisioning as an explicit and measurable deployment decision for LLM agents in mission\-critical infrastructure\. A task taxonomy derived from the mathematical formulation of physical systems characterizes task demand, while a cumulativeK1K\_\{1\}–K5K\_\{5\}hierarchy characterizes the information and capabilities available to the agent\. We estimate the relationship between them using a literature\-derived map of prior provisioning practices and an execution\-derived map that selects the lowest harness level achieving near\-best class\-level performance\. These maps guide either direct lookup or self\-check\-triggered escalation toK5K\_\{5\}\. Our results show that suitable harness configurations vary across task classes and domains, and that additional provision may improve, leave unchanged, or reduce performance\. They support selecting harness configurations according to task and domain requirements rather than granting maximal information, tools, and operational access by default\.

Limitations and Future Work\.The measured task\-to\-harness relationship is calibrated to the evaluated executor, harness implementation, protocol, and tolerance\. Its generalization to new domains, models, and harness implementations requires further validation and, where necessary, recalibration\. Our evaluation covers 240 tasks in two simulation\-backed environments, which do not fully capture live conditions such as sensor uncertainty, distribution shift, organizational procedures, and human approval\. The literature\-derived map may reflect publication and annotation bias, while the execution\-derived map remains subject to finite\-sample variation, judge error, and imperfect self\-checks\. Future work will expand and release a broader benchmark for LLM agents in MCI O&M, together with reproducible test environments covering additional domains, tasks, disturbances, and tool interfaces\. We will also investigate learning\-based methods that use the task\-to\-harness map to guide harness selection and escalation, and extend the evaluation framework to include security and operational risks\.

## References

- Aggarwalet al\.\(2024\)P\. Aggarwal, A\. Madaan, A\. Anand, S\. P\. Potharaju, S\. Mishra, P\. Zhou, A\. Gupta, D\. Rajagopal, K\. Kappaganthu, Y\. Yang, S\. Upadhyay, M\. Faruqui, and MausamAutoMix: automatically mixing language models\.InProceedings of the 38th International Conference on Neural Information Processing Systems,NIPS ’24,Red Hook, NY, USA\.External Links:ISBN 9798331314385Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1),[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p3.1)\.
- Asaiet al\.\(2024\)A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. HajishirziSelf\-rag: learning to retrieve, generate, and critique through self\-reflection\.InInternational conference on learning representations,Vol\.2024,pp\. 9112–9141\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1)\.
- \[3\]Cybersecurity and Infrastructure Security AgencyCritical infrastructure security and resilience\.Note:[https://www\.cisa\.gov/topics/critical\-infrastructure\-security\-and\-resilience](https://www.cisa.gov/topics/critical-infrastructure-security-and-resilience)Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p1.1)\.
- Duet al\.\(2024\)Y\. Du, F\. Wei, and H\. ZhangAnyTool: self\-reflective, hierarchical agents for large\-scale api calls\.InProceedings of the 41st International Conference on Machine Learning,ICML’24\.Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1)\.
- Endsley \(1995\)M\. R\. EndsleyToward a theory of situation awareness in dynamic systems\.Human Factors: The Journal of the Human Factors and Ergonomics Society37\(1\),pp\. 32–64\.Cited by:[§4\.2](https://arxiv.org/html/2608.17433#S4.SS2.p1.1)\.
- Franko \(2025\)U\. FrankoDynamic system instructions and tool exposure for efficient agentic llms\.arXiv preprint arXiv:2602\.17046\.Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1),[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p3.1)\.
- Google DeepMind \(2026\)Google DeepMindGemini 3\.1 Pro: Model Card\.Note:[https://deepmind\.google/models/model\-cards/gemini\-3\-1\-pro/](https://deepmind.google/models/model-cards/gemini-3-1-pro/)Cited by:[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p4.1)\.
- Huet al\.\(2025\)M\. Hu, P\. Zhao, C\. Xu, Q\. Sun, J\. Lou, Q\. Lin, P\. Luo, and S\. RajmohanAgentGen: enhancing planning abilities for large language model based agent via environment and task generation\.InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.1,KDD ’25,New York, NY, USA,pp\. 496–507\.External Links:ISBN 9798400712456,[Link](https://doi.org/10.1145/3690624.3709321),[Document](https://dx.doi.org/10.1145/3690624.3709321)Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p1.1)\.
- Jianget al\.\(2023\)Z\. Jiang, F\. Xu, L\. Gao, Z\. Sun, Q\. Liu, J\. Dwivedi\-Yu, Y\. Yang, J\. Callan, and G\. NeubigActive retrieval augmented generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 7969–7992\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.495/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.495)Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1)\.
- Jorenet al\.\(2025\)H\. Joren, J\. Zhang, C\. Ferng, D\. Juan, A\. Taly, and C\. RashtchianSufficient context: a new lens on retrieval augmented generation systems\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 20310–20334\.Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1)\.
- Kimet al\.\(2025\)J\. Kim, W\. Choi, and B\. LeePrompt flow integrity to prevent privilege escalation in llm agents\.arXiv preprint arXiv:2503\.15547\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1),[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.
- Laiet al\.\(2025\)S\. Lai, Z\. Xu, W\. Zhang, H\. Liu, and H\. XiongLLMLight: large language models as traffic signal control agents\.InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.1,KDD ’25,New York, NY, USA,pp\. 2335–2346\.External Links:ISBN 9798400712456,[Link](https://doi.org/10.1145/3690624.3709379),[Document](https://dx.doi.org/10.1145/3690624.3709379)Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p2.1)\.
- Laptevet al\.\(2015\)N\. Laptev, S\. Amizadeh, and I\. FlintGeneric and scalable framework for automated time\-series anomaly detection\.InProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,pp\. 1939–1947\.Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p1.1)\.
- Leeet al\.\(2026\)Y\. Lee, R\. Nair, Q\. Zhang, K\. Lee, O\. Khattab, and C\. FinnMeta\-harness: end\-to\-end optimization of model harnesses\.arXiv preprint arXiv:2603\.28052\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p1.1)\.
- Liet al\.\(2026\)J\. Li, X\. Xiao, Y\. Zhang, C\. Liu, L\. Zhao, X\. Liao, Y\. Ji, J\. Wang, Y\. Ge, W\. Xu, X\. Fang, X\. Xu, T\. Zhao, Y\. Kim, J\. Hamm, T\. Wang, and C\. ReddyAgent harness engineering: a survey\.External Links:[Link](https://openreview.net/pdf?id=eONq7FdiHa)Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p1.1),[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p1.1)\.
- Liet al\.\(2022\)M\. Li, Z\. Li, K\. Yin, X\. Nie, W\. Zhang, K\. Sui, and D\. PeiCausal inference\-based root cause analysis for online service systems with intervention recognition\.InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,KDD ’22,New York, NY, USA,pp\. 3230–3240\.External Links:ISBN 9781450393850,[Link](https://doi.org/10.1145/3534678.3539041),[Document](https://dx.doi.org/10.1145/3534678.3539041)Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p1.1)\.
- Liuet al\.\(2026\)C\. Liu, Y\. Guo, Y\. Liu, Y\. Yang, Q\. Yan, X\. Zhao, W\. Hua, S\. Liu, S\. Li, Y\. Bu,et al\.Auditing agent harness safety\.arXiv preprint arXiv:2605\.14271\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.
- Luet al\.\(2023\)P\. Lu, B\. Peng, H\. Cheng, M\. Galley, K\. Chang, Y\. N\. Wu, S\. Zhu, and J\. GaoChameleon: plug\-and\-play compositional reasoning with large language models\.InProceedings of the 37th International Conference on Neural Information Processing Systems,NIPS ’23,Red Hook, NY, USA\.Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1)\.
- Maet al\.\(2025a\)M\. Ma, J\. Clark, and S\. ZhangAIOpsLab in action: an open platform for aiops research\.InProceedings of the 33rd ACM International Conference on the Foundations of Software Engineering,FSE Companion ’25,New York, NY, USA,pp\. 1223–1227\.External Links:ISBN 9798400712760,[Link](https://doi.org/10.1145/3696630.3728619),[Document](https://dx.doi.org/10.1145/3696630.3728619)Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p2.1)\.
- Maet al\.\(2025b\)Z\. Ma, J\. Liu, X\. Luo, Z\. Huang, Q\. Zhu, and W\. CheAdvancing tool\-augmented large language models via meta\-verification and reflection learning\.InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2,KDD ’25,New York, NY, USA,pp\. 2078–2089\.External Links:ISBN 9798400714542,[Link](https://doi.org/10.1145/3711896.3736835),[Document](https://dx.doi.org/10.1145/3711896.3736835)Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p1.1)\.
- Marotet al\.\(2021\)A\. Marot, B\. Donnot, G\. Dulac\-Arnold, A\. Kelly, A\. O’Sullivan, J\. Viebahn, M\. Awad, I\. Guyon, P\. Panciatici, and C\. RomeroLearning to run a power network challenge: a retrospective analysis\.InNeurIPS 2020 competition and demonstration track,pp\. 112–132\.Cited by:[§5\.3](https://arxiv.org/html/2608.17433#S5.SS3.p1.1)\.
- Menget al\.\(2026\)Q\. Meng, Y\. Wang, L\. Chen, Y\. Li, W\. Wu, W\. Jiang, Q\. Wang, C\. Lu, Y\. Gao, Y\. Wu,et al\.Agent harness for large language model agents: a survey\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p1.1),[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p1.1)\.
- Mohammadiet al\.\(2025\)M\. Mohammadi, Y\. Li, J\. Lo, and W\. YipEvaluation and benchmarking of llm agents: a survey\.InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2,KDD ’25,New York, NY, USA,pp\. 6129–6139\.External Links:ISBN 9798400714542,[Link](https://doi.org/10.1145/3711896.3736570),[Document](https://dx.doi.org/10.1145/3711896.3736570)Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p2.1)\.
- Mudunuriet al\.\(2026\)S\. Mudunuri, J\. Wan, A\. Qin, and S\. ManoharanSemantic tool discovery for large language models: a vector\-based approach to mcp tool selection\.arXiv preprint arXiv:2603\.20313\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1)\.
- Onget al\.\(2024\)I\. Ong, A\. Almahairi, V\. Wu, W\. Chiang, T\. Wu, J\. E\. Gonzalez, M\. W\. Kadous, and I\. StoicaRouteLLM: learning to route llms with preference data\.External Links:2406\.18665,[Link](https://arxiv.org/abs/2406.18665)Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1)\.
- OpenAI \(2026\)OpenAIIntroducing GPT‑5\.4\.Note:[https://openai\.com/index/introducing\-gpt\-5\-4/](https://openai.com/index/introducing-gpt-5-4/)Cited by:[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p2.1)\.
- Panet al\.\(2026\)L\. Pan, L\. Zou, S\. Guo, J\. Ni, and H\. ZhengNatural\-language agent harnesses\.arXiv preprint arXiv:2603\.25723\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p1.1)\.
- Patelet al\.\(2026\)D\. Patel, C\. Shyalika, S\. R\. Yarrabothula, L\. Yue, S\. Lin, N\. Zhou, and J\. RayfieldResults and retrospective analysis of the cods 2025 assetopsbench challenge\.arXiv preprint arXiv:2605\.08518\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p2.1)\.
- Qinet al\.\(2024\)Y\. Qin, S\. Liang, Y\. Ye, K\. Zhu, L\. Yan, Y\. Lu, Y\. Lin, X\. Cong, X\. Tang, B\. Qian,et al\.Toolllm: facilitating large language models to master 16000\+ real\-world apis\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 9695–9717\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p2.1)\.
- Rasmussen \(1985\)J\. RasmussenThe role of hierarchical knowledge representation in decisionmaking and system management\.IEEE Transactions on systems, man, and cybernetics\(2\),pp\. 234–243\.Cited by:[§4\.2](https://arxiv.org/html/2608.17433#S4.SS2.p1.1)\.
- Renet al\.\(2019\)H\. Ren, B\. Xu, Y\. Wang, C\. Yi, C\. Huang, X\. Kou, T\. Xing, M\. Yang, J\. Tong, and Q\. ZhangTime\-series anomaly detection service at microsoft\.InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,KDD ’19,New York, NY, USA,pp\. 3009–3017\.External Links:ISBN 9781450362016,[Link](https://doi.org/10.1145/3292500.3330680),[Document](https://dx.doi.org/10.1145/3292500.3330680)Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p1.1)\.
- Shiet al\.\(2025\)T\. Shi, J\. He, Z\. Wang, L\. Wu, H\. Li, W\. Guo, and D\. SongProgent: programmable privilege control for llm agents\.arXiv e\-prints,pp\. arXiv–2504\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1),[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.
- Shinnet al\.\(2023\)N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. YaoReflexion: language agents with verbal reinforcement learning\.InProceedings of the 37th International Conference on Neural Information Processing Systems,NIPS ’23,Red Hook, NY, USA\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p5.1),[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p1.1),[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p3.1)\.
- Suet al\.\(2019\)Y\. Su, Y\. Zhao, C\. Niu, R\. Liu, W\. Sun, and D\. PeiRobust anomaly detection for multivariate time series through stochastic recurrent neural network\.InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,KDD ’19,New York, NY, USA,pp\. 2828–2837\.External Links:ISBN 9781450362016,[Link](https://doi.org/10.1145/3292500.3330672),[Document](https://dx.doi.org/10.1145/3292500.3330672)Cited by:[§2\.1](https://arxiv.org/html/2608.17433#S2.SS1.p1.1)\.
- Thurneret al\.\(2018\)L\. Thurner, A\. Scheidler, F\. Schafer, J\. H\. Menke, J\. Dollichon, F\. Meier, S\. Meinecke, and M\. BraunPandapower \- an open source python tool for convenient modeling, analysis and optimization of electric power systems\.IEEE Transactions on Power Systems\.External Links:[Document](https://dx.doi.org/10.1109/TPWRS.2018.2829021),[Link](https://arxiv.org/abs/1709.06743),ISSN 0885\-8950Cited by:[§5\.3](https://arxiv.org/html/2608.17433#S5.SS3.p1.1)\.
- Yanget al\.\(2026\)K\. Yang, Y\. Bu, J\. Yi, Y\. Wang, B\. Zhou, J\. Dai, S\. Hu, and Y\. YangWhen lower privileges suffice: investigating over\-privileged tool selection in llm agents\.arXiv preprint arXiv:2606\.20023\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.
- Yaoet al\.\(2023\)S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. CaoReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p1.1)\.
- Yaoet al\.\(2026\)Y\. Yao, X\. Tan, C\. Liu, Y\. Li, Z\. Wang, W\. Yu, Z\. Tan, Y\. Tian, G\. Zhao, L\. Sun,et al\.Harness\-bench: measuring harness effects across models in realistic agent workflows\.arXiv preprint arXiv:2605\.27922\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p1.1)\.
- Yuet al\.\(2026\)J\. Yu, J\. Zhu, B\. Lin, Q\. Cui, Z\. Ding, and X\. LiSkill is not one\-size\-fits\-all: model\-aware skill alignment for llm agents\.arXiv preprint arXiv:2605\.30723\.Cited by:[§1](https://arxiv.org/html/2608.17433#S1.p2.1)\.
- Zhanget al\.\(2025\)K\. Zhang, Z\. Su, P\. Chen, E\. Bertino, X\. Zhang, and N\. LiLLM agents should employ security principles\.arXiv preprint arXiv:2505\.24019\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.
- Zhaoet al\.\(2024\)A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. HuangExpeL: llm agents are experiential learners\.InProceedings of the Thirty\-Eighth AAAI Conference on Artificial Intelligence and Thirty\-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence,AAAI’24/IAAI’24/EAAI’24\.External Links:ISBN 978\-1\-57735\-887\-9,[Link](https://doi.org/10.1609/aaai.v38i17.29936),[Document](https://dx.doi.org/10.1609/aaai.v38i17.29936)Cited by:[§2\.2](https://arxiv.org/html/2608.17433#S2.SS2.p1.1),[§7\.1](https://arxiv.org/html/2608.17433#S7.SS1.p3.1)\.
- Zhuet al\.\(2025\)J\. Zhu, K\. Tseng, G\. Vernik, X\. Huang, S\. G\. Patil, V\. Fang, and R\. A\. PopaMiniScope: a least privilege framework for authorizing tool calling agents\.arXiv preprint arXiv:2512\.11147\.Cited by:[§2\.3](https://arxiv.org/html/2608.17433#S2.SS3.p2.1)\.

## Appendix ALiterature Corpus and Annotation

### A\.1\.Corpus Collection and Screening

We survey how the last decade \(2016\-\-2026\) of MCI operation\-and\-maintenance research provisions system access, covering 13 of the 16 CISA critical\-infrastructure sectors999[https://www\.cisa\.gov/topics/critical\-infrastructure\-security\-and\-resilience/critical\-infrastructure\-sectors](https://www.cisa.gov/topics/critical-infrastructure-security-and-resilience/critical-infrastructure-sectors)\(Defense Industrial Base, Financial Services, and Government Facilities are out of scope\)\. Candidates come from three public indices, arXiv, Semantic Scholar, and OpenAlex, queried with sector vocabulary only \(“substation”, “chiller plant”, …\): no task or method words, so the task and harness distributions emerge from the papers rather than from the query set\. The collection and screening pipeline was built and operated with Claude Code \(Fable 5\)\. From roughly 2,000 crawled candidates, an LLM screening agent triages title and abstract, discarding surveys, position papers, and papers without a primary MCI O&M task, and full\-text extraction runs on the survivors\. The screened corpus contains 1,223 papers, of which 1,220 receive complete task\-coordinate and harness\-level annotations and are used to estimate the literature\-derived map; Figure[5](https://arxiv.org/html/2608.17433#A1.F5)shows its sector distribution and vocabulary\.

![Refer to caption](https://arxiv.org/html/2608.17433v1/figures/term_wordcloud_final.png)

Figure 5\.The literature corpus\. Left: sector distribution\. Right: data\-driven task and method vocabulary; word size is corpus frequency\.
### A\.2\.Task and Harness Annotation Schema

Each paper receives two annotations from separate model calls that never see each other’s output, a task coordinate and a harness level, so the class label cannot leak into the harness label\. The task call is method\- and name\-blind: it returns the paper’s primary task as one method\-free sentence plus only the three axis labels of the taxonomy; no class names appear anywhere in the prompt\. The harness call assigns the highest level the proposed solution*materially*uses, considering both what the deployed method consumes at run time and what building it required, a policy trained in a simulator is K5 even if inference reads only telemetry\. Figure[6](https://arxiv.org/html/2608.17433#A1.F6)summarizes both rubrics in the paper’s vocabulary\.

Task class⟨ω,τ,e⟩\\langle\\omega,\\tau,e\\rangle\. *Output mode*ω\\omega—Inform: report information about the target;Act: select an intervention intended to affect the target\. *Target time*τ\\tau—Now\(tt\): the current system;Future\(t\+Δt\+\\Delta\): an outcome at a later time\. *Target element*ee— observablesignalsyy; latentstatesxx; systemmechanismsmm\. Harness levelKK\(highest level materially used\)\. K1K\_\{1\}: Model\-only reasoning\.The method uses only the task input and model parametric knowledge;K2K\_\{2\}: Static knowledge\.\+\+manuals, SOPs, specifications, design documents, or rule bases;K3K\_\{3\}: Temporal observations\.\+\+telemetry, logs, alarms, or time\-series measurements;K4K\_\{4\}: Structure and physics\.\+\+topology, component relations, governing equations, physical constraints, or control logic;K5K\_\{5\}: Forward simulation\.\+\+executable simulation, digital\-twin rollouts, counterfactual evaluation, or optimization over hypothetical future trajectories\.

Figure 6\.Abridged annotation rubric aligned with the task and harness spaces defined in Section[4](https://arxiv.org/html/2608.17433#S4)\.
### A\.3\.Multi\-LLM Annotation Protocol

We use three annotators from different model providers, includinggemini\-3\.1\-pro,gpt\-5\.4\-mini, andglm\-5\.2, to independently assign each paper a task coordinate⟨ω,τ,e⟩\\langle\\omega,\\tau,e\\rangleand a harness levelKK, together with supporting evidence from the paper\. The task axes are consolidated separately by majority vote, while the harness level is consolidated as a single ordinal label\. Majority voting directly resolves 89% of task\-axis labels and 97% of harness labels\. The remaining cases, comprising 30 task\-coordinate assignments and 36 harness\-level assignments, undergo a separate adjudication pass assisted by Claude Fable 5\. The adjudicator reviews the paper evidence without access to the original model votes, and the resulting labels are manually verified\. Table[4](https://arxiv.org/html/2608.17433#A1.T4)reports the consensus composition and inter\-annotator agreement\.

We will release the complete extraction through an interactive literature explorer, including each paper’s task coordinate, harness level, supporting evidence, individual annotations, and consolidated labels\. The explorer also supports inspection and filtering across task classes and harness levels; Figure[7](https://arxiv.org/html/2608.17433#A1.F7)shows a snapshot\.

Table 4\.Consensus composition and inter\-annotator agreement\.![Refer to caption](https://arxiv.org/html/2608.17433v1/figures/fig_paper_explorer.png)Figure 7\.Snapshot of the interactive literature explorer to be released with the project\.Each point represents one paper, grouped by task class and consensus harness level\. The inspection panel presents the selected paper’s metadata, supporting evidence, individual annotations, and consolidated labels\.
### A\.4\.Corpus\-Level Findings

Figure[2](https://arxiv.org/html/2608.17433#S5.F2)shows the harness distribution for each task class, while Table[5](https://arxiv.org/html/2608.17433#A1.T5)reports the counts used to estimatePlit​\(K∣T\)P\_\{\\mathrm\{lit\}\}\(K\\mid T\)\. Each task axis is associated with higher harness provision: the mean level increases by0\.450\.45from Inform to Act and by0\.390\.39from Now to Future, while the marginal mean rises from3\.533\.53for observable signals to3\.653\.65for system mechanisms and4\.064\.06for latent states\. Provision is concentrated atK3K\_\{3\}andK5K\_\{5\}, which account for53%53\\%and28%28\\%of the corpus, respectively, compared with13%13\\%atK4K\_\{4\}\.

Figure[8](https://arxiv.org/html/2608.17433#A1.F8)provides a joint view of these trends over the complete2×2×32\\times 2\\times 3task space\. Harness provision generally increases toward Act, Future, and latent\-state tasks, with the highest values concentrated around A\-F\-xxand adjacent Future–Act classes\. The variation is nevertheless not determined by any single axis: task classes sharing one or two coordinates can still exhibit different literature\-derived harness distributions\.

The corpus records the harness level used by prior methods rather than the level sufficient for a fixed agent and environment\. Estimates for sparsely represented classes, particularly A\-N\-xx\(n=5n=5\) and A\-F\-xx\(n=11n=11\), are also less reliable\. We therefore treat the literature\-derived map as a prior rather than verified execution evidence\. More broadly, the analysis provides a survey\-scale account of how MCI O&M research has provisioned system information and capabilities across task classes\.

![Refer to caption](https://arxiv.org/html/2608.17433v1/figures/literature_3d_landscape.png)Figure 8\.Three\-dimensional view of the literature\-derived harness landscape\.The vertices represent the 12 task classes defined by output mode, target time, and target element, and surface colour indicates the mean literature\-derived harness levelK¯\\bar\{K\}\. Conventional task names are shown only as approximate interpretations of the formal coordinates\.Table 5\.Literature\-derived harness distributions across the 12 task classes, based on the consensus labels of three annotators\. Counts are reported by harness level, and task classes are ordered by meanKK\.

## Appendix BBenchmark Construction and Verification

### B\.1\.Simulation Environments

Liquid\.A differentiable thermal\-hydraulic digital twin of a 10\-rack liquid\-cooled data hall: one CDU loop \(pump, control valve, facility heat exchanger\) feeding 20 server cold plates over a 130\-node/149\-edge hydraulic network\. Transient states integrate withdopri5atΔ​t=5\\Delta t\{=\}5s \(rtol10−410^\{\-4\}, atol10−610^\{\-6\}\); steady states solve the coupled hydraulic–thermal system directly\. An episode spans 0–1200 s\. Telemetry is a∼\\sim910\-column table offamily:targetsignals \(node temperatures and pressures, flows, pump speed, per\-plate power\); the internal ODE states \(e\.g\., metal base\-plate temperatures\) are recorded but hidden from the agent, which grounds the latent\-state \(xx\) task classes\.

Grid\.A power\-grid environment on Grid2Op’sl2rpn\_case14\_sandbox\(IEEE\-14: 14 substations, 20 branches, 6 generators, 11 loads\), with pandapower AC power flow as the K5 engine\. Reference trajectories are 288\-step days at 5\-minute resolution with overflow disconnection disabled so that overloads remain observable\. Telemetry is a 142\-column table \(per\-line loadingρ\\rho, flows, currents, voltages, status; per\-generator and per\-load injections\)\. Here latent\-state tasks arise from*structurally absent*columns—network losses and the full bus\-voltage vector are in no column, and three buses have no voltage sensor—rather than from masking\.

### B\.2\.Scenario Generation

Liquid uses six simulated trajectories \(two normal\-operation trajectories, single\- and multi\-plate overheating, a migrating hotspot, and a pump degradation ramp\)\. Grid uses ten one\-day windows selected by a two\-pass scan over 1,004 Grid2Op chronics: two normal days, seven sustained\-overload windows spanning three distinct bottleneck lines at increasing severity \(ρmax\\rho\_\{\\max\}1\.00–1\.26\), and one mid\-window line\-trip fault that cascades toρ=1\.95\\rho\{=\}1\.95on a neighboring line\. Every task binds a scenario file, an observation window, and a current time; the visibility gate \(Appendix[C](https://arxiv.org/html/2608.17433#A3)\) prevents future\-peeking, so scenarios are replayable and deterministic\.

### B\.3\.Task Construction

Each domain instantiates 10 tasks for each of the 12 classes \(240 total\) from parameterized template families \(e\.g\.,value\_at,hidden\_plate\_temp,best\_action\_rho,min\_pump\_below\_maxtemp\), with natural\-language paraphrase applied on top of templates\. Every task carries a machine\-readable ground\-truth specificationgt\_spec=\{method, args\} resolved by a deterministic oracle \(∼\\sim80 methods for Liquid, 54 for Grid\) with three evidence tiers: direct trajectory lookups, closed\-form physics \(e\.g\.,Q=ρ​V˙​cp​Δ​TQ=\\rho\\dot\{V\}c\_\{p\}\\Delta T\), and simulator co\-solves that call the*same*engine exposed to the agent at K5, with privileged access \(full trajectory, no visibility gate, hidden states\)\. Benchmark templates are written against the system representation, not against any literature paper, so the literature map and the benchmark share only the taxonomy\.

### B\.4\.Construction and Test Splits

Each class splits 5/5 into a construction \(map\-estimation\) split and a held\-out test split, stratified by template family so no family appears only in one split, with a fixed RNG seed\. The fullK1K\_\{1\}–K5K\_\{5\}sweep runs on all 240 tasks, but execution\-derived levels are estimated*only*from construction tasks and all policies are evaluated*only*on test tasks\.

### B\.5\.Programmatic and Human Verification

Programmatic checks: \(1\) the ground truth, re\-fed verbatim to the grader, must score 1\.0 on all 240 tasks; \(2\) a cross\-split echo audit verifies that no test answer string appears in construction traces; \(3\) independent oracle re\-derivation validates every stored ground truth; \(4\) a probe pass removes degenerate axes discovered during construction \(e\.g\., the slack generator’s inert setpoint and a voltage\-pinned bus in Grid, which would otherwise make some “counterfactual” tasks trivial or dead\)\. Human review of sampled traces additionally corrected two Grid ground\-truth values after an audit of the prediction/prognosis cells; the affected 20 tasks were re\-executed at all levels before any map or policy result reported here\.

### B\.6\.Representative Tasks

Table[6](https://arxiv.org/html/2608.17433#A2.T6)shows one test task per class from the Liquid domain \(Grid analogues replace temperatures with line loadings and pump/valve actions with generation redispatch\)\.

Table 6\.Representative Liquid benchmark tasks, one per class\. S/X/M denote signal\(yy\)/state\(xx\)/mechanism targets\(mm\); graders are deterministic rule checks combined with an LLM judge \(Appendix[D](https://arxiv.org/html/2608.17433#A4)\)\.

## Appendix CHarness Implementations

### C\.1\.Cumulative Harness Manifests

Levels are strictly cumulative \(ℋK1⊂⋯⊂ℋK5\\mathcal\{H\}\_\{K\_\{1\}\}\\subset\\cdots\\subset\\mathcal\{H\}\_\{K\_\{5\}\}\): each level’s registry unions the previous level’s tools, and an agent at levelKKis constructed with exactly that registry\. Table[7](https://arxiv.org/html/2608.17433#A3.T7)lists both manifests\.

Table 7\.Resources exposed at each harness level\. Tools are OpenAI function\-calling interfaces; each level unions all lower levels\.
### C\.2\.Tool and Data Interfaces

All tools are exposed as JSON function\-calling schemas; results return as structured JSON \(time series are capped at 100 points per call, transient rollouts at 18 steps\)\. K5 override conventions are uniform \(pump speed, valve opening, per\-plate power in Liquid; generator/load MW and line outages in Grid\)\. Simulator wall\-time and call counts are metered separately per task\.

### C\.3\.Capability Enforcement

Capability boundaries are architectural, not prompt\-level: a level\-KKagent’s process is constructed with only that level’s tool registry, so higher\-level tools are not merely discouraged but nonexistent in its API schema\. A shared visibility gate clamps every data access to the task’s observation horizon \(agent\-supplied times are clamped server\-side, including K5 rollout start times\), and hidden state families are masked at the data source; the grading oracle deliberately bypasses both\. During escalation, only the model’s own textual findings are carried into the next attempt—no tool outputs, caches, or context cross the boundary\.

### C\.4\.Domain\-Specific Differences

K1–K3 are semantically identical across domains \(same tool names, samekind:targetsignal convention, same agent shell\)\. K4 differs by physics content only \(hydraulic network and heat balances vs\. electrical topology, thermal limits, and power balances\)\. K5 differs structurally: Liquid exposes a transient rollout because minute\-scale thermal dynamics are genuine physics \(plate thermal mass, exchanger lag\), whereas the grid is quasi\-static at 5\-minute resolution—given injections, the state follows algebraically from power flow—so steady AC power flow plus N\-1 screening*is*the complete simulation capability\. A grid time\-rollout is deliberately not exposed: the grid’s future is driven by exogenous injection trajectories, so stepping the environment forward would either replay recorded future injections \(leaking ground truth through the visibility gate\) or reduce to the already\-provided counterfactual solve under hypothesized injections; extrapolating those injections from telemetry is the agent’s own K3\-level task, not a harness capability\. Latent state is gated telemetry in Liquid but structural sensor absence in Grid—two realistic mechanisms for the same taxonomy coordinate\.

## Appendix DExecution and Evaluation Protocol

### D\.1\.Executor Configuration

All main results use a frozengpt\-5\.4executor behind a ReAct\-style function\-calling loop: one model call per iteration, emitted tool calls executed and returned, until the model answers without tool calls or hits the iteration cap \(20\)\. Sampling: temperature 0\.2, no output\-token cap\. Each task runs in an isolated spawned process with a 300 s hard timeout and a soft deadline 45 s earlier; on exhaustion the agent receives one final tool\-free “answer from the evidence you have” call, so timeouts degrade to a best\-effort answer rather than an empty one\. The cross\-executor study servesQwen3\.5\-27Bon a single\-node vLLM server \(thinking disabled, native tool\-call parser, server\-side prompt\-tail truncation to fit the context window\) behind the identical loop, harness, and tasks\.

### D\.2\.Task Scoring

Each task is scored in\[0,1\]\[0,1\]as the mean of a deterministic rule score and an LLM\-judge score\. Rule graders are typed per task: numeric\-with\-tolerance \(key\-anchored extraction with unit normalization\), set match \(Jaccard\), boolean polarity, directional, and required\-fact containment\. Inform tasks grade the reported estimate against the oracle value; Act tasks grade the selected intervention against the simulator\-verified best action \(e\.g\., the cheapest pump setting that satisfies the constraint\), so action quality is grounded in the same physics the oracle uses\.

### D\.3\.LLM Judge

The judge isgemini\-3\.1\-proat temperature 0 with a fixed rubric \(1\.0 fully correct→\\to0\.0 wrong or empty; credit substance over phrasing and unit differences\)\. It sees only the question, ground truth, and final answer—never the harness level, policy, tool trace, or any experience text—so it cannot systematically favor a provisioning condition\.

### D\.4\.Cost Accounting

Token counts sum the API\-reported usage of every executor\-side LLM call in a condition—routing calls, self\-checks, every escalation attempt, and verification calls for cascades—so policies are charged for their full mechanism; judge tokens are metered separately and never count toward policy cost\. Latency is wall\-clock per task, summed across attempts, and includes tool and simulator execution\. Because absolute costs depend on the serving stack, the main text reports both normalized to Full\-K5K\_\{5\}within each domain; absolute reference values appear in Appendix[G](https://arxiv.org/html/2608.17433#A7)\.

### D\.5\.Repetition and Randomness

Every condition runs three independent repetitions; tables report the mean±\\pmpopulation std of the three run means\. Scenarios are deterministic replays, so nondeterminism enters only through model sampling\. Execution\-derived levels are estimated from the pooled construction\-split runs; policies are evaluated on the disjoint test split\.

## Appendix EMap Estimation and Statistical Analysis

### E\.1\.Literature\-Derived Map

We setAlit​\(T,K\)=Plit​\(K∣T\)A\_\{\\mathrm\{lit\}\}\(T,K\)=P\_\{\\mathrm\{lit\}\}\(K\\mid T\)andϵlit=0\.05\\epsilon\_\{\\mathrm\{lit\}\}=0\.05\. Under this tolerance,𝒢lit\\mathcal\{G\}\_\{\\mathrm\{lit\}\}selects the modal harness level for all 12 task classes\.

### E\.2\.Execution\-Derived Map

For each task classTT,μexec​\(T,K\)\\mu\_\{\\mathrm\{exec\}\}\(T,K\)is the mean score over the five construction tasks and three runs\. The execution\-derived map selects the lowest harness level whose score is withinϵexec=0\.05\\epsilon\_\{\\mathrm\{exec\}\}=0\.05of the best observed level\. Table[9](https://arxiv.org/html/2608.17433#A5.T9)reports the selected levels and their stability\.

### E\.3\.Tolerance Sensitivity

Table[8](https://arxiv.org/html/2608.17433#A5.T8)sweepsϵexec\\epsilon\_\{\\text\{exec\}\}\. Grid floors are invariant forϵ≤0\.05\\epsilon\\leq 0\.05and Liquid changes only two classes between 0 and 0\.05; test accuracy is flat within noise across 0\.05–0\.15 while token cost varies<<10%\. The paper’s operating pointϵ=0\.05\\epsilon\{=\}0\.05sits on this plateau, so no conclusion depends on the tolerance choice\.

Table 8\.Sensitivity of the execution\-derived map toϵexec\\epsilon\_\{\\text\{exec\}\}\. “Δ\\Deltafloors” counts classes whose selected level differs fromϵ=0\.05\\epsilon\{=\}0\.05; accuracy/tokens are test\-split lookup results\.
### E\.4\.Execution\-derived Level Stability

Table[9](https://arxiv.org/html/2608.17433#A5.T9)resamples the five construction tasks per class \(1,000 bootstrap draws\) and reports the probability of re\-selecting the adopted floor, plus the set of floors reachable by leave\-one\-task\-out\. In 19 of 24 class×\\timesdomain cells the adopted floor is re\-selected with probability≥0\.7\\geq 0\.7; instability concentrates where adjacent levels are within tolerance of each other \(e\.g\., Liquid I\-F\-S, where K3/K4/K5 all solve the class\), i\.e\., exactly where the choice is least consequential\. No resampling ever moves a floor below K2\.

Table 9\.Execution\-derived level stability atϵ=0\.05\\epsilon\{=\}0\.05: bootstrap probability of re\-selecting the adopted floor over construction tasks, and leave\-one\-out floor sets\.
### E\.5\.Statistical Tests

Paired bootstrap over the 60 test tasks \(per\-task scores averaged over three runs;10410^\{4\}resamples\): on Liquid, Exec\-Lookup vs\. Full\-K5K\_\{5\}givesΔ=\+0\.018\\Delta\{=\}\{\+\}0\.018, 95% CI\[−0\.022,\+0\.061\]\[\-0\.022,\+0\.061\]—statistically indistinguishable accuracy at 14% lower token cost, which is the claimed trade\. On Grid, Exec\-Lookup is genuinely below Full\-K5K\_\{5\}\(Δ=−0\.044\\Delta\{=\}\{\-\}0\.044, CI\[−0\.087,−0\.008\]\[\-0\.087,\-0\.008\]\) and genuinely above Lit\-Lookup \(Δ=\+0\.098\\Delta\{=\}\{\+\}0\.098, CI\[\+0\.020,\+0\.183\]\[\+0\.020,\+0\.183\]\), confirming both that maximal provision stays accuracy\-optimal on Grid and that execution evidence is needed to calibrate the literature prior\.

### E\.6\.Oracle Definitions

The Class Oracle selects, per class, the level with the highest test\-split mean; the Task Oracle selects the best level per individual task; both use mean scores across the three runs\. Both peek at test outcomes and are therefore non\-deployable ceilings; neither has a consistent cost, so they appear as accuracy lines only\.

## Appendix FBaseline Implementations

### F\.1\.Full\-K5K\_\{5\}

The frozen executor with the complete K5 registry on every task; no routing calls\.

### F\.2\.Instruction–Tool Retrieval

The official ITR package with default configuration \(hybrid dense \+ BM25 \+ cross\-encoder retrieval\)\. The tool corpus is the 14–16 harness tool specs verbatim; the instruction corpus is the K5 system prompt in fragments, with exemplars from successful construction traces \(Grid\)\. Per task, retrieved tools map to the minimal cumulative level containing them \(retrieval confidence<0\.7<0\.7falls back to full provision\); routing itself costs zero LLM tokens\.

### F\.3\.LLM\-Route and LLM\+Exp

One tool\-free routing call to the same executor, presented with a generic legend of the five levels and instructed to pick the cheapest sufficient level, terminating inLEVEL: Kx\. LLM\+Exp additionally shows the router a distilled experience playbook: per\-class recipe cards distilled by the executor from construction\-split trajectories \(multi\-run scores, best tool sequences with observation snippets, contrastive failures; the pre\-registered floor annotations are withheld\), synthesized into a≤\\leq1,200\-word global playbook\. The playbook informs routing only; execution runs on the raw task\. Routing tokens count toward cost\.

### F\.4\.AutoMix

The official AutoMix implementation mapped from a model cascade to a harness cascade: the K3 episode plays the small model, the K5 episode the large one\. Confidence uses the official few\-shot self\-verification prompt at temperature 1\.0 withk=8k\{=\}8samples; routing uses the official 8\-bin POMDP meta\-verifier trained on construction\-split rows \(threshold fallback\), with per\-domain costs set to measured mean tokens\. Verification tokens count toward cost, which is why AutoMix is expensive despite reusing episodes\.

### F\.5\.Blind\-ESC

Identical loop, self\-check protocol, one\-tier\-per\-failure climbing, finding\-carrying, and grading as Map\-ESC; the only difference is the seed \(K1K\_\{1\}vs\. the mapped floor\)\. It therefore isolates the value of the map from the value of escalation\.

### F\.6\.Reflexion and ExpeL

Reflexion runs at fixed K5 with the official reflection prompts and at most three attempts; since benchmark rewards are hidden at run time, the retry trigger is a ground\-truth\-free LLM self\-verdict \(PASS/FAIL\), and reflections read the real tool trace\. ExpeL runs at fixed K5 with the official pipeline: 20 induced rules from construction trajectories plus 6 task\-similar few\-shot exemplars retrieved by sentence\-embedding kNN from 50 successful construction traces\.

## Appendix GFull Experimental Results

### G\.1\.AggregateK1K\_\{1\}–K5K\_\{5\}Sweep

Table[10](https://arxiv.org/html/2608.17433#A7.T10)gives the full\-corpus sweep \(all 240 tasks×\\times3 runs\)\. Aggregate performance is not strictly monotonic, but the aggregate is exactly what the per\-class analysis shows to be misleading: In Liquid,K4K\_\{4\}performs belowK3K\_\{3\}in aggregate despite improving performance for several individual task classes\. In Grid,K5K\_\{5\}is the most accurate level and uses fewer tokens thanK3K\_\{3\}andK4K\_\{4\}, although it is not the least costly level overall\.

Table 10\.FullK1K\_\{1\}–K5K\_\{5\}sweep, all 120 tasks per domain, 3 runs\.
### G\.2\.Per\-Class Execution Scores

Table[11](https://arxiv.org/html/2608.17433#A7.T11)reports the mean accuracy for all ten tasks in each class underK1K\_\{1\}–K5K\_\{5\}\. These full\-benchmark results complement, rather than reproduce, the construction\-split scores shown in Figure[3](https://arxiv.org/html/2608.17433#S5.F3)\.

Table 11\.Per\-class mean accuracy±\\pmstd for every harness level \(10 tasks/class, 3 runs\)\.
### G\.3\.Construction\- and Test\-Split Analysis

Figure[3](https://arxiv.org/html/2608.17433#S5.F3)reports construction\-split scores, from which the execution\-derived map is estimated\. All provisioning policies are evaluated on the disjoint test split in Table[1](https://arxiv.org/html/2608.17433#S7.T1)\. Re\-estimating the map on the test split changes the selected level in 10 of the 24 class–domain pairs, primarily in the classes identified as less stable in Table[9](https://arxiv.org/html/2608.17433#A5.T9)\. Using the construction\-split map results in a mean class\-level test regret of0\.0360\.036, consistent with the gap between Exec\-Lookup and the Class Oracle\.

### G\.4\.Absolute Reference Costs

Normalization anchors for Table 1 \(test split, Full\-K5K\_\{5\}\): Liquid16,21016\{,\}210tokens and47\.247\.2s per task; Grid13,40913\{,\}409tokens and7\.67\.6s per task\. Multiplying Table 1 ratios by these anchors recovers absolute costs; e\.g\., Exec\-Lookup averages≈14\.0\{\\approx\}14\.0k tokens/3939s on Liquid and11,82911\{,\}829tokens/8\.68\.6s on Grid\.

## Appendix HAdditional Analyses

### H\.1\.Routing Granularity and Construction Bias

The class is the right routing unit for this corpus: family\-level routing \(floors per template family\) reaches 0\.634/0\.704 \(Liquid/Grid\) and task\-level nearest\-neighbor routing 0\.639/0\.690, both below class\-level lookup \(0\.670/0\.762\)\. Preferred harnesses also vary*across*families within 7 of 12 Liquid and 5 of 12 Grid classes, so class\-level floors are not an artifact of any single template family; they smooth over family\-level sampling noise that finer\-grained routing overfits\.

### H\.2\.Escalation Diagnostics

Map\-ESC escalates rarely: 11\.7% of Liquid and 13\.9% of Grid test executions trigger theK5K\_\{5\}retry \(mean 1\.12/1\.14 attempts per task\)\. Seeds follow the map \(Liquid: 105/180 executions start atK5K\_\{5\}, 60 atK3K\_\{3\}, and 15 atK2K\_\{2\}\.\), and most below\-K5 seeds finish where they started \(Liquid: 39 of 60 K3\-seeded executions; Grid: 20 of 45\)—the self\-check acts as tail\-risk insurance on a calibrated start, not as a router\.

*Step\-size ablation: is the directK5K\_\{5\}jump too coarse?*We ran the gradual alternative—Ki\+1=min⁡\(Ki\+1,K5\)K\_\{i\+1\}=\\min\(K\_\{i\}\{\+\}1,K\_\{5\}\), findings carried at every hop—for three repetitions in both domains\. It is dominated: accuracy falls to0\.692±0\.0090\.692\\pm 0\.009\(Liquid\) and0\.762±0\.0100\.762\\pm 0\.010\(Grid\) versus0\.715±0\.0140\.715\\pm 0\.014and0\.782±0\.0100\.782\\pm 0\.010for the direct jump, at equal cost on Liquid and higher cost on Grid \(\+5%\+5\\%tokens,\+8%\+8\\%latency\)\. The gradual paths explain why: of the K3\-seeded executions that escalated, only 4 of 19 \(Liquid\) and 1 of 24 \(Grid\) were satisfied by the intermediate K4 step—the rest climbed on toK5K\_\{5\}anyway, having spent an extra attempt at still\-insufficient provision and carried a failed intermediate trace into the final context\. The insufficiency that survives a correctly\-mapped seed is rarely one tier deep, so the intermediate grant buys little; escalating directly to full provision is simultaneously more accurate and more efficient, which is why Map\-ESC uses the two\-point design\.

### H\.3\.Alternative Experience Representations

Consistent with the main text, execution experience helps most as a*measurement*\(the map\) rather than as text: injecting the distilled playbook into execution at Full\-K5K\_\{5\}helps Grid \(0\.810 vs\. 0\.806\) but not Liquid \(0\.648 vs\. 0\.652\), and routing\-only use of the playbook \(LLM\+Exp\) does not beat the free lookup in either domain\. ExpeL’s trajectory retrieval is the strongest execution\-side use of experience on Liquid \(0\.731\) but costs 44% more tokens than Map\-ESC for a comparable gain\.

### H\.4\.Qualitative Error Analysis

Four recurring failure modes: \(i\)*under\-provisioning*: below the floor, the agent either abstains or extrapolates from typical ranges \(Liquid I\-N\-S at K1/K2 scores 0\.00—the honest failure the taxonomy predicts\); \(ii\)*over\-provisioning backfire*: with simulation available, agents run counterfactuals instead of reading the trend \(Liquid I\-N\-X drops from 0\.59 at K3 to 0\.46 at K5; Grid A\-N\-X from 0\.69 to 0\.65\), and sensitivity sweeps burn the iteration budget; \(iii\)*executor\-limited classes*: Liquid A\-F\-X never exceeds 0\.35 at any level—transient counterfactual deltas exceed what the executor can orchestrate, and no provisioning policy can fix this; \(iv\)*self\-check errors*: false passes dominate false escalations, which is why Blind\-ESC stalls below K5 on classes whose confident\-but\-wrong answers never trigger the retry\.

## Appendix IArtifact Availability and Responsible Deployment

### I\.1\.Released Artifacts

We will release both 120\-task corpora with ground\-truth specifications and splits, the frozen scenario telemetry \(CSV\), the harness manifests and tool schemas for all five levels, the literature corpus IDs with all three per\-model annotations and adjudications, the distilled experience playbooks, evaluation and map\-estimation code, and per\-task execution records \(scores, token/latency/simulator meters, tool\-call traces\) sufficient to reproduce every table and figure without API access\.

### I\.2\.Proprietary Components

The liquid\-cooling twin is currently proprietary\. We will release its frozen trajectories, topology metadata, and recorded tool I/O, which are sufficient to reproduce all reported results\. Generating new Liquid trajectories currently requires access to the twin; once it is open\-sourced, we will release the complete Liquid environment\. The Grid environment, based on Grid2Op and pandapower, is fully open\-source\.

### I\.3\.Compute and API Usage

The twoK1K\_\{1\}–K5K\_\{5\}sweeps consumed 17\.1M \(Liquid\) and 17\.7M \(Grid\) executor tokens plus∼\\sim2\.4M judge tokens; policy and baseline conditions add a comparable amount, for∼\\sim75M tokens total across all reported experiments\. Qwen experiments ran on a single\-node vLLM deployment; simulator compute is negligible \(sub\-second per solve, metered per task\)\.