Paper deep dive
Evo-Harness: Context-to-Harness Skill Compilation for Self-Evolving Agents
Tianxin Wei, Zhan Shi, Minhua Lin, Bing He, Zewen Liu, Yisi Sang, Yuanchen Bei, Xuying Ning, Jiaru Zou, Ting-Wei Li, Xiao Lin, Yanjun Zhao, Chi Wang, Benoit Dumoulin, Dakuo Wang, Jingrui He, Hanqing Lu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/18/2026, 5:23:05 AM
Summary
The paper introduces Evo-Harness, a framework for online harness learning where a frozen Large Language Model (LLM) agent improves by incrementally updating a structured external skill harness across sequential tasks. The core mechanism, context-to-harness skill compilation, distills noisy, single-shot execution contexts (particularly failures) into reusable general and topic-level skills. Evaluated on five benchmarks (TerminalBench2, SWE-bench, CL-Bench, Tau-bench, WebArena-Infinity), Evo-Harness demonstrates superior performance compared to baselines like AWM, Dynamic Cheatsheet, and XSkill, highlighting the importance of compiling experiences rather than just retrieving them.
Entities (13)
Relation Signals (11)
Online Harness Learning → isformulatedby → Evo-Harness
confidence 95% · To address these gaps, we formulate online harness learning... through our proposed Evo-Harness.
Evo-Harness → uses → Context-to-Harness Skill Compilation
confidence 95% · At its core, context-to-harness skill compilation distills noisy, single-shot executions into reusable skill harnesses
Evo-Harness → evaluatedon → SWE-BENCH
confidence 92% · evaluate across five realistic benchmarks (...SWE-bench...)
Evo-Harness → evaluatedon → CL-Bench
confidence 92% · evaluate across five realistic benchmarks (...CL-Bench...)
Evo-Harness → evaluatedon → tau-bench
confidence 92% · evaluate across five realistic benchmarks (...τ-bench...)
Evo-Harness → evaluatedon → WebArena-Infinity
confidence 92% · evaluate across five realistic benchmarks (...WebArena-Infinity).
Evo-Harness → evaluatedon → TerminalBench2
confidence 92% · evaluate across five realistic benchmarks (TerminalBench2...)
Evo-Harness → outperforms → AWM
confidence 90% · EVO-HARNESS improves over both No-Evolve and prior experience-based baselines across all five benchmarks.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Learning from experience is critical for developing capable, self-improving large language model (LLM) agents. Existing methods typically extract knowledge from accumulated trajectories via reflection, memory, rules, or skills. However, agents in realistic environments continuously encounter novel tasks, often offering only a one-shot opportunity to improve. These executions yield rich but highly noisy contexts, entangling broadly useful lessons with task-specific artifacts. Critically, prior works rarely validate their effectiveness on complex real-world tasks or isolate the underlying drivers of improvement. To address these gaps, we formulate online harness learning, where a frozen agent improves by continually updating a structured harness across sequential tasks. This formulation enables a systematic study of key self-improvement factors through our proposed Evo-Harness. At its core, context-to-harness skill compilation distills noisy, single-shot executions into reusable skill harnesses for cross-domain and topic-level adaptation. To demonstrate the efficacy of one-shot skill compilation, we evaluate across five realistic benchmarks (TerminalBench2, SWE-bench, CL-Bench, -bench, WebArena-Infinity). Our extensive analysis demonstrates the effectiveness of Evo-Harness and provides a principled understanding of how LLM agents can effectively learn on the fly. Our code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.15071v1
- Canonical: https://arxiv.org/abs/2608.15071v1
Trouble viewing inline? Open PDF directly →
Full Text
59,932 characters extracted from source content.
Expand or collapse full text
EVO-HARNESS: Context-to-Harness Skill Compilation for Self-Evolving Agents Tianxin Wei 1 , Zhan Shi 2 , Minhua Lin 3 , Bing He 2 , Zewen Liu 4 , Yisi Sang 2 , Yuanchen Bei 1 , Xuying Ning 1 , Jiaru Zou 1 , Ting-Wei Li 1 , Xiao Lin 1 , Yanjun Zhao 1 , Chi Wang 5 , Benoit Dumoulin 2 , Dakuo Wang 6 , Jingrui He 1 , Hanqing Lu 2 1 University of Illinois Urbana-Champaign 2 Amazon 3 The Pennsylvania State University 4 Emory University 5 AG2 AI 6 Northeastern University Abstract Learning from experience is critical for devel- oping capable, self-improving large language model (LLM) agents. Existing methods typi- cally extract knowledge from accumulated tra- jectories via reflection, memory, rules, or skills. However, agents in realistic environments con- tinuously encounter novel tasks, often offering only a one-shot opportunity to improve. These executions yield rich but highly noisy contexts, entangling broadly useful lessons with task- specific artifacts. Critically, prior works rarely validate their effectiveness on complex real- world tasks or isolate the underlying drivers of improvement. To address these gaps, we for- mulate online harness learning, where a frozen agent improves by continually updating a struc- tured harness across sequential tasks. This formulation enables a systematic study of key self-improvement factors through our proposed Evo-Harness. At its core, context-to-harness skill compilation distills noisy, single-shot exe- cutions into reusable skill harnesses for cross- domain and topic-level adaptation. To demon- strate the efficacy of one-shot skill compilation, we evaluate across five realistic benchmarks (TerminalBench2, SWE-bench, CL-Bench,τ- bench, WebArena-Infinity). Our extensive anal- ysis demonstrates the effectiveness of EVO- HARNESS and provides a principled under- standing of how LLM agents can effectively learn on the fly. Our code is available at Link. 1 Introduction Large language model (LLM) agents (Achiam et al., 2023; Team et al., 2023; Liu et al., 2024) have shown strong performance in complex task- solving settings, such as web interaction (Zhou et al., 2024), software engineering (Jimenez et al., 2024), tool use (Qin et al., 2024), and long-horizon reasoning (Erdogan et al., 2025; Sun et al., 2025). These settings require agents to plan, act, interpret observations, handle feedback, and recover from errors. However, when an agent fails on a task, the Past Experiences Retrieve Focus: recalling past experiences or strategies Context Compile Guide Focus: Harness and compile skill guidance Past Experiences Skill Harness Apply (similar tasks) Guide (new contexts) Filter Case Bank (a) Experience Reuse (b) Harness Skill Evolution Figure 1: Comparison of experience reuse (retrieving past cases) and harness skill evolution (compiling expe- riences into reusable guidance). failure often remains an isolated event rather than a source of future improvement. Without convert- ing failed executions into reusable lessons, agents continue to make similar mistakes in later tasks. Many experience-based methods (Shinn et al., 2023; Zhao et al., 2024) address this issue by learn- ing from accumulated trajectories offline (Lee et al., 2026; Ni et al., 2026), utilizing media such as re- flection (Shinn et al., 2023; Qiu et al., 2026), mem- ory (Zhong et al., 2024; Chhikara et al., 2025; Xu et al., 2025), and skills (Zhang et al., 2026a; Xia et al., 2026; Jiang et al., 2026; Yang et al., 2026). However, in realistic settings, agents continuously encounter novel tasks and contexts, and each com- pleted execution typically offers only a one-shot op- portunity to improve future behavior. Furthermore, real executions produce rich but highly noisy con- texts, entangling useful lessons with task-specific details, partial failures, and incidental tool traces. Critically, prior methods often fail to rigorously validate their effectiveness under these realistic con- straints, nor do they systematically isolate the un- derlying drivers of agent improvement. To navigate arXiv:2608.15071v1 [cs.AI] 15 Aug 2026 (a) Offline Experience-based Learning Collect Experience ...... ... ... Analyze & Mine Pattern ... Skill 1Skill 2 Skill M ... ... (b) Practical Online Task Stream IterateSkillsOffline ... Experience 2 Content Actions Outcome Issues ... Experience 3 Content Actions Outcome Issues ... Experience N Content Actions Outcome Issues ... Task 1Task 2 Task 3 Task N Experience 1 Content Actions Outcome Issues ... Time (c) Evo-Harness: Online Skill Compilation Harness 푯 ퟏ Harness 푯 ퟐ Harness 푯 ퟑ Harness 푯 ퟒ Compile memories from each Update harness skills Guide subsequent tasks Figure 2: Illustration of the problem setting. (a) Prior methods collect executions and mine reusable patterns offline. (b) In practical online streams, agents encounter tasks sequentially, yielding rich but one-shot execution experiences. (c) EVO-HARNESS compiles these online experiences into evolving skill harnesses for future reuse. such a one-shot, complex learning environment, re- lying on the simple retrieval of past experiences falls short. Instead, agents need to actively compile raw execution contexts into reusable skill guidance for novel tasks, as contrasted in Figure 1. Building on this view, we formulate online har- ness learning, where a frozen agent improves by incrementally updating a structured, external har- ness across sequential tasks. To cope with the con- tinuous influx of novel tasks, this harness acts as a self-evolving learning medium that progressively distills executions into actionable guidance for the future. An overview of this practical setting, com- pared to standard offline learning, is illustrated in Figure 2. To operationalize this formulation and enable a systematic study of continuous self-improvement, we introduce Evo-Harness. We instantiate the har- ness within this framework using skills. Skills act as a natural medium for realistic task executions because they can integrate diverse learning signals, such as failures, feedback, constraints, and opera- tional procedures, into structured knowledge that shapes future planning and recovery. At the core of Evo-Harness is context-to-harness skill compila- tion, which converts individual execution contexts into reusable skill harness updates. Specifically, our compilation process uses a solver to reflect on con- texts and propose candidate skills, and an evolver to update the existing harness. General skills capture transferable cross-task patterns, while topic skills capture localized operational knowledge. This dual design preserves critical learning signals while fil- tering out trajectory-level noise. We evaluate Evo-Harness across five realistic benchmarks (TerminalBench2, SWE-bench, CL- Bench,τ-bench, and WebArena-Infinity) under a sequential continuous learning setup. Beyond demonstrating performance gains, our primary fo- cus is systematically investigating the core drivers of self-improvement. By analyzing how evolver designs, feedback types, and transfer settings dic- tate performance, our study treats the skill harness as an interpretable object for understanding how frozen LLM agents learn on the fly. Our contributions are summarized as follows: • Online harness learning. We formulate a setting where frozen agents improve over sequential task streams by incrementally updating a reusable ex- ternal harness instead of model parameters. • Context-to-harness skill compilation. We in- troduce a mechanism that compiles single-shot execution contexts into actionable general and topic-level skill updates. • Systematic analysis of harness evolution. We isolate how evolver designs, feedback types, and transfer settings impact continuous self- improvement. Evaluated across five complex benchmarks, EVO-HARNESS consistently out- performs existing experience-based methods. 2 Related Work In this section, we review existing works on self- evolving LLM agents and external memory or skill harnesses for agent experience. Evolving and self-improving LLM agents. A growing line of work explores how LLM agents self-improve through interaction and feedback. Be- yond local reflection methods that iteratively revise outputs based on scalar or verbal feedback (Shinn et al., 2023; Madaan et al., 2023), experience- based approaches attempt to extract reusable knowl- edge offline. For instance, some works mine ac- cumulated trajectory pools to synthesize natural language insights (Zhao et al., 2024), interaction rules (Chen et al., 2024), or transferable directo- ries (Ni et al., 2026). Similarly, a parallel track focuses on passively evolving higher-level agent artifacts, such as prompts (Suzgun et al., 2025; Lou et al., 2026), reasoning banks (Ouyang et al., 2025), or memory structures (Agrawal et al., 2025; Lee et al., 2026; Zhang et al., 2025; Wei et al., 2025; Lin et al., 2026a). To provide more operational guidance, recent methods have shifted toward skill- centered evolution. This includes building exe- cutable code skills (Wang et al., 2023), optimiz- ing skill libraries via policy learning (Xia et al., 2026; Zhang et al., 2026a; Ouyang et al., 2026), or curating skill knowledge bases and representa- tions (Yang et al., 2026; Qiu et al., 2026; Jiang et al., 2026; Wang et al., 2026). While closely related, these existing frameworks largely treat learned arti- facts as passive repositories for free-form retrieval. Critically, prior methods often fail to rigorously validate their effectiveness under realistic online constraints, nor do they systematically isolate the underlying drivers of agent improvement. Our work addresses this gap by compiling single-shot experiences into an active skill harness, providing a structural scaffold to systematically investigate how LLM agents learn on the fly. Memory and skill harnesses for agent experi- ence. Memory provides a fundamental external medium for preserving agent experience. Existing memory systems maintain long-term observation histories (Zhong et al., 2024; Yu et al., 2025), or- ganize memories through adaptive structures (Xu et al., 2025), and explore methods for updating and retrieving evidence across modalities (Huang et al., 2026; Bei et al., 2026; Liu et al., 2026a). How- ever, these approaches largely treat memory as a passive information store. Skills differ fundamen- tally because they encode operational knowledge. They shift from merely recording what happened to specifying exactly how an agent should act through manuals (Chen et al., 2024), verify operations via organized directories (Ni et al., 2026), or recover from failures using skill repositories (Ouyang et al., 2026; Wang et al., 2026; Zhang et al., 2026b). This procedural view elevates skills into an active har- ness. Rather than merely supplying historical con- text, a harness acts as an external scaffold that struc- turally shapes future planning, taking forms such as pattern repositories (Ning et al., 2026; Qiu et al., 2026), reasoning banks (Ouyang et al., 2025), and dynamic guidance files (Suzgun et al., 2025; Lee et al., 2026; Liu et al., 2026b; Lin et al., 2026b). Building on this distinction, our work conceptual- izes the skill harness not merely as a performance- enhancing artifact, but as an interpretable medium to systematically study continuous online adapta- tion. To navigate the high noise of real-world ex- ecutions, EVO-HARNESS compiles single-shot re- flections into a structured harness of general and topic skills. While this separation efficiently filters out task-specific artifacts, its primary purpose in our work is to provide a transparent scaffold. This explicitly allows us to systematically isolate and evaluate the core drivers, such as evolver designs and feedback types, that dictate how frozen LLM agents self-improve on the fly. 3EVO-HARNESS: Context-to-Harness Skill Compilation We present EVO-HARNESS, a framework for on- line harness learning. The goal is to convert one- shot execution contexts into a reusable harness that guides future task solving. Unlike experience re- trieval, which recalls past cases, EVO-HARNESS compiles noisy execution signals into structured guidance that can shape later planning, acting, veri- fication, and recovery. Figure 3 gives an overview. EVO-HARNESS maintains a current harness that guides the solver before execution, shaping how it plans, acts, ver- ifies, and recovers. After each task batch, execu- tion contexts are reflected into candidate memories and compiled back into the harness as cross-task patterns and task-type procedures, allowing the up- dated harness to guide later executions. 3.1 Online Harness Learning Setup LetB 1 ,...,B K denote a stream of task batches, whereB i = x i,1 ,...,x i,m i . Before batchB i , the agent has a current harnessH i , a structured set of reusable guidance entries: H i =h 1 i ,...,h n i i .(1) The harness is external to the frozen solver and serves as a control scaffold for later executions. Unlike raw trajectory storage, it maintains compact guidance distilled from prior executions and affects future behavior through selection and injection into the solver context. For each taskx i,j , the agent selects a compact set of harness entries: S i,j = Select(x i,j ,H i ;b), |S i,j |≤ b,(2) Collect Experience Task Batch / Stream ... Current Skill Harness Solver Agent Execution Context input actions outcome issues ... Exec 1 ... Execution Reflection Memory 1 (from Exec 1) Memory 2 (from Exec 2) Memory N (from Exec N) ... ... Task 1 Task 2 Task N ... ... Evolver Compile & Manage Reflected Memories Current Harness Cross-task Patterns Task-type Procedures Continue evolving harness input actions outcome issues ... Exec 2 input actions outcome issues ... Exec N Candidates Compare Edit Figure 3: Overview of EVO-HARNESS. The solver uses the current harness to solve a task batch or stream. Each execution produces a context, including input, action trajectory, outcome, and feedback. Execution reflection converts these contexts into candidate memories. The evolver compares the candidate memories with the current harness and updates the harness with cross-task patterns and task-type procedures for later batches. wherebis the injection budget. The selected guid- ance is injected into the task: ̃x i,j = Inject(x i,j ,S i,j ).(3) A frozen solverA executes the task: (τ i,j ,y i,j ,f i,j ) =A( ̃x i,j ),(4) whereτ i,j is the action trajectory,y i,j is the out- come, andf i,j is available feedback, such as ver- ifier results, unit-test outputs, tool diagnostics, or judge feedback. The execution context is: c i,j = (x i,j ,τ i,j ,y i,j ,f i,j ).(5) Online harness learning updatesH i intoH i+1 from these execution contexts, while keeping the solver parameters fixed. 3.2 Context-to-Harness Compilation EVO-HARNESS uses two complementary stages: execution reflection and harness evolution. Reflec- tion extracts learning signals from each execution. Evolution integrates those signals into the harness. Execution reflection.For each completed execu- tion contextc i,j , the solver produces a candidate memory only when the execution fails or receives negative feedback: r i,j = Reflect(c i,j ).(6) Each candidate memory is represented as: r = (lesson, trigger, evidence, scope_hint). (7) Reflection uses the full task-side context avail- able immediately after execution, including the instruction, trajectory, observations, tool outputs, outcome, and feedback. We focus reflection on failures because they expose the solver’s current boundary: incorrect assumptions, missing con- straints, ineffective tool use, weak verification, and failed recovery. This makes the resulting candidate memories more targeted for harness improvement, while avoiding unnecessary updates from success- ful executions that may contain task-specific details. The reflection stage does not compare against the current harness, so its output is not a final harness update, but a failure-grounded candidate signal ex- tracted from the completed execution. Harness evolution.After batchB i , the candidate memories form: R i =r i,j m i j=1 .(8) The evolver receives the candidate memories and the current harness: O i = Evolver(H i ,R i ),(9) whereO i is a set of harness edits. Conceptually, the evolver applies: o = π evolve (r,H i ,R i ),(10) whereo ∈ ADD, MERGE, REVISE, SKIP. The evolver filters noisy or redundant memories, merges compatible signals, revises existing guid- ance, and promotes reusable lessons when appro- priate. The updated harness is: H i+1 = ApplyEdits(H i ,O i ).(11) This makes the harness more than a memory store: candidate memories are not appended directly, but compiled into guidance that can affect future be- havior. Algorithm 1 EVO-HARNESS: Context-to-Harness Skill Compilation 1:Input: task batchesB 1 ,...,B K , frozen solverA, initial harnessH 1 2: for i = 1,...,K do 3: R i ←∅ 4:for each task x i,j ∈ B i do 5: S i,j ← Select(x i,j ,H i ;b) 6: ̃x i,j ← Inject(x i,j ,S i,j ) 7:(τ i,j ,y i,j ,f i,j )←A( ̃x i,j ) 8:c i,j ← (x i,j ,τ i,j ,y i,j ,f i,j ) 9:r i,j ← Reflect(c i,j ) 10: R i ←R i ∪r i,j 11:end for 12: O type i ← CompileTaskType(H i ,R i ) 13: O cross i ← CompileCrossTask(H i ,R i ) 14: O i ←O type i ∪O cross i 15: H i+1 ← ApplyEdits(H i ,O i ) 16: end for 17: Output: evolved harnessH K+1 and task results 3.3 Harness Levels The evolved harness contains two levels of reusable guidance. Cross-task patterns are compiled by comparing memories across tasks and identifying shared operational structure. Task-type procedures preserve more localized guidance for recurring task formats, interfaces, or domains. These levels are not predefined labels for individual executions. They emerge from how the evolver filters, com- pares, and integrates candidate memories into the harness. 3.4 Algorithm and Modularity Algorithm 1 summarizes EVO-HARNESS. We use batch-level updates so the evolver can both compile localized task-type procedures and compare memo- ries across tasks to extract cross-task patterns. The same formulation reduces to task-type updating when|B i | = 1. Here,CompileTaskTypepreserves localized procedures, whileCompileCrossTaskextracts shared patterns across executions. The modular design is intentional: by decoupling the solver, evolver, feedback source, and initial harness state, EVO-HARNESS provides a controlled framework for systematically studying what drives online har- ness improvement. 4 Experiments We conduct experiments to evaluate EVO- HARNESS and to understand when online harness evolution improves LLM agents: •RQ1: Overall effectiveness.Does EVO- HARNESS improve frozen agents across diverse benchmarks, and which task categories benefit most? •RQ2: Harness content and components. What types of guidance are produced by the evolved harness, and how do different harness compo- nents contribute? •RQ3: Online updating and transfer. How do evolved harnesses behave under train-split trans- fer, online updating, and solver-evolver pairing? •RQ4: Feedback grounding. How does the source and granularity of feedback affect harness evolution? 4.1 Setup Experimental setup.We evaluate EVO- HARNESS on five benchmarks spanning diverse agent capabilities: WebArena-Infinity (Zhou, 2026) for web navigation, TerminalBench2 (Merrill et al., 2026) for command-line tasks, SWE-bench Lite (Jimenez et al., 2024) for software engineering, CL-Bench (Dou et al., 2026) for adaptive reason- ing, and TAU-Bench (Yao et al., 2024) for tool- use tasks. We use Claude Opus 4.6 (Anthropic, 2026a) as the default solver and evolver, and addi- tionally evaluate with Claude Opus 4.7 (Anthropic, 2026b), Claude Opus 4.5 (Anthropic, 2025), Kimi- K2.5 (Team et al., 2026), and GPT-OSS (Agarwal et al., 2025) when studying model ablations. We report success rates in percent, where higher is bet- ter. Further benchmark and implementation details are provided in Appendix A and F. Baselines.We compare with representative experience-, memory-, and skill-based adapta- tion methods under the same task split and eval- uation protocol.AWM (Wang et al., 2024) reuses agent working memory, Dynamic Cheat- sheet (Suzgun et al., 2025) maintains an adap- tive external cheatsheet, Evo-Memory (Wei et al., 2025) evolves reusable memories from executions, ACE (Zhang et al., 2025) optimizes agent context, and XSkill (Jiang et al., 2026) separates experience and skills for continual reuse. DC-Cu and DC-RS are two Dynamic Cheatsheet variants. 4.2 Overall Effectiveness (RQ1) Table 1 reports the main results with Claude Opus 4.6 as the solver.EVO-HARNESS improves over both No-Evolve and prior experience-based baselines across all five benchmarks. The gains are especially large on TerminalBench-2, where agents need to inspect files, execute commands, in- Table 1: Main results using Claude Opus 4.6 as the solver. Scores denote success rates (%), and higher values indicate better performance. EVO-HARNESS achieves the best results across all five benchmarks. MethodCL-BenchTerminalBench-2SWE-bench Lite τ -benchWebArena-Infinity No Evolve29.5462.9263.6772.7372.50 AWM28.9162.9262.6770.9170.00 DC-Cu29.2360.6763.0072.1271.25 DC-RS29.0761.8061.6770.9170.00 Evo-Memory29.3864.0464.0072.1272.50 ACE29.7061.8063.6772.7371.25 XSkill31.4466.2964.6773.9473.75 EVO-HARNESS34.0273.0367.0076.9776.25 Figure 4: Distribution of evolved skill patterns across benchmarks. The learned skills exhibit benchmark- specific characteristics, ranging from workflow and tool- use procedures to verification, domain reasoning, and output-constraint handling. terpret errors, and recover from failed operations. This suggests that harness evolution is particularly useful when tasks expose reusable procedural struc- ture. Compared with the strongest baseline, XSkill, EVO-HARNESS still improves on every benchmark. External experience reuse is not automatically beneficial. Several baselines underperform No- Evolve on some benchmarks, suggesting that re- trieved or evolved guidance can be noisy, overly specific, or mismatched to the solver. This supports our motivation for studying not only whether self- evolution helps, but also what factors make online harness updates effective. 4.3 Category and Model Generality (RQ1) Table 2 provides a representative category-level view on CL-Bench across five solver models. Since CL-Bench covers diverse reasoning and execution categories, it allows us to examine where harness evolution is most effective and how the effect varies across solver capabilities. EVO-HARNESS im- proves the overall score for every solver model, with larger gains on stronger Claude models than on open-weight alternatives. Opus 4.7, Opus 4.6, and Opus 4.5 obtain gains of+3.7,+4.5, and+3.8, respectively, while Kimi-K2.5 and GPT- OSS improve by+1.1and+0.8. This suggests that stronger solvers may be better able to interpret, follow, and benefit from evolved harness guidance. Across categories, the most consistent improve- ments appear on Procedural Task Execution (PTE), where every model improves and Opus 4.6 and Opus 4.5 each gain+9.8points. Harness evolu- tion is most reliable when tasks expose reusable operational structure. In contrast, Empirical Dis- covery & Simulation (EDS) shows less stable gains, with drops for Opus 4.6 and GPT-OSS. This sug- gests that more exploratory tasks may be more sensitive to overly specific guidance. 4.4 Harness Content and Components (RQ2) Figure 4 analyzes the content of the evolved harness across benchmarks. The evolved har- ness reflects benchmark-specific operational de- mands rather than accumulating generic guid- ance. WebArena-Infinity mainly yields workflow and procedure skills, TerminalBench-2 emphasizes verification and recovery,τ-bench mixes tool/API operation with domain reasoning, and CL-Bench produces a broader distribution including domain reasoning and output-constraint handling. These patterns suggest that context-to-harness compila- tion adapts the harness to each environment’s oper- ational structure and failure modes. Table 3 studies which parts of the harness up- date process contribute to performance. We focus this and the following diagnostic analyses mainly on CL-Bench and SWE-bench Lite because they stress different forms of harness guidance: CL- Bench contains heterogeneous adaptive reasoning tasks where localized task-type guidance can mat- ter, while SWE-bench Lite emphasizes repository- level debugging and verification procedures where broader reusable guidance is expected to help. NO Table 2: CL-Bench category results across different models: pass rate (%) across five models. Bold = best per category-model pair.∆shows absolute improvement of EVO-HARNESS over the No-Evolve baseline. DKR = Domain Knowledge Reasoning, EDS = Empirical Discovery & Simulation, PTE = Procedural Task Execution, RSA = Rule System Application. DKREDSPTERSAAll ModelMethod Rate∆Rate∆Rate∆Rate∆Rate∆ Opus 4.7 No-Evolve 32.4—19.1—35.5—29.0—30.8— EVO-HARNESS 34.1(+1.7)19.6(+0.5)42.7(+7.2)32.7(+3.7)34.5(+3.7) Opus 4.6 No-Evolve 31.4—20.6—34.0—26.9—29.5— EVO-HARNESS 35.6(+4.2)18.1(-2.5)43.8(+9.8)30.4(+3.5)34.0(+4.5) Opus 4.5 No-Evolve31.4—18.1—27.8—25.3—27.3— EVO-HARNESS32.6(+1.2)19.1(+1.0)37.6(+9.8)28.3(+3.0)31.1(+3.8) Kimi-K2.5 No-Evolve25.6—12.1—25.1—21.6—22.9— EVO-HARNESS25.9(+0.3)13.1(+1.0)29.3(+4.2)20.8(-0.7)23.9(+1.1) GPT-OSS No-Evolve17.6—11.6—17.6—16.4—16.6— EVO-HARNESS17.9(+0.3)10.6(-1.0)18.5(+0.8)18.6(+2.1)17.5(+0.8) Table 3: Ablation study on CL-Bench and SWE-bench Lite using Claude Opus 4.6 as the solver. Scores denote pass rates (%). MethodCL-BenchSWE-bench Lite No Evolve29.5463.67 No Propose33.2865.33 General Only30.2866.67 Topic Only33.7064.33 EVO-HARNESS34.0267.00 PROPOSE removes solver-side proposal of candi- date memories, GENERAL ONLY keeps only cross- task guidance, and TOPIC ONLY keeps only local- ized task-type guidance. Both broad cross-task guidance and local- ized task-type guidance are useful, but their importance depends on the task setting. The full EVO-HARNESS setting performs best on both benchmarks. On CL-Bench, Topic Only reaches 33.70, close to the full result of34.02, while Gen- eral Only drops to30.28, suggesting that localized task-type guidance is more useful for its heteroge- neous reasoning tasks. On SWE-bench Lite, Gen- eral Only reaches66.67, close to the full result of 67.00, while Topic Only reaches64.33, indicating that broad debugging, verification, and repository- inspection procedures matter more. No Propose improves over No Evolve but remains below the full method, suggesting that proposal and harness- level compilation are complementary. 4.5 Transfer and Adaptability (RQ3) Figure 5 compares three skill-update settings. NO- EVOLVE uses no harness update. TRAIN-SPLIT Figure 5: SWE-bench Lite success rates under dif- ferent evolution settings. NO-EVOLVE uses no har- ness updates; TRAIN-SPLIT TRANSFER evaluates skills evolved by Claude Sonnet on the training split; and ONLINE-UPDATED updates skills during the task stream. TRANSFER uses skills evolved by Claude Sonnet 4.5 on the training split and evaluates them with an Opus 4.7 solver on the test split. ONLINE- UPDATED updates the harness directly with the same models during the test task stream. Skills learned on a training split can transfer across both tasks and models, while online updating performs best. No-Evolve reaches68.8, Train- Split Transfer reaches73.4, and Online-Updated reaches75.0. The gain from Train-Split Transfer shows that skills evolved by a smaller or differ- ent model can still provide useful guidance for a stronger solver on unseen tasks. Online-Updated further improves over Train-Split Transfer by+1.6 points, suggesting that in-situ harness updates bet- ter adapt to the local distribution and failure modes of the test stream. Figure 6 examines how solver-evolver pairing affects harness evolution. We compare two solver (a) Opus 4.7 Solver(b) Sonnet 4.5 Solver Figure 6: Effect of solver-evolver pairing on SWE- bench Lite. SAME denotes using the same model for solving and skill evolution, while CROSS denotes using different models for the solver and evolver. settings: an Opus 4.7 solver and a Sonnet 4.5 solver. In each setting, SAME uses the same model as both solver and evolver, while CROSS uses the other model as the evolver. Specifically, for the Opus solver, SAME uses Opus to evolve the harness and CROSS uses Sonnet to evolve the harness; for the Sonnet solver, SAME uses Sonnet as the evolver and CROSS uses Opus as the evolver. Using the same model for solving and evolution is not al- ways optimal; cross-model evolution can help, but only when the solver is capable of using the evolved guidance. For the Opus solver, Cross reaches76.0, slightly higher than Same at75.3and clearly above No-Evolve at70.7. This suggests that a different evolver can produce guidance that is still useful, or even more useful, when the down- stream solver has sufficient capability to interpret and apply it. For the Sonnet solver, however, Same and Cross reach55.3and55.7, both below No- Evolve at58.0. Thus, Sonnet-evolved skills can help Opus, but do not necessarily help Sonnet itself. This contrast suggests that harness transfer depends not only on the skill artifact but also on the base solver’s capability: if the solver cannot reliably fol- low, contextualize, or adapt the guidance, evolution may fail to improve performance. 4.6 Feedback Grounding (RQ4) Table 4 evaluates how feedback signals affect harness evolution on CL-Bench and SWE-bench Lite. We compare three feedback settings. SELF- GENERATED asks the LLM itself to judge whether the execution succeeded and what should be learned. MINIMAL provides only the environment- level success or failure signal, such as whether a command, verifier, or test passed. STANDARD provides grounded diagnostic feedback, including failure messages, error traces, verifier outputs, or test failures when available. Table 4: Impact of feedback level on CL-Bench and SWE-bench Lite using Claude Opus 4.6 as the solver. Scores denote pass rates (%). Feedback LevelCL-BenchSWE-bench Lite No Evolve29.5463.67 Self-Generated27.9661.67 Minimal29.8667.33 Standard34.0267.00 Grounded feedback is important for reliable harness evolution. Self-Generated feedback un- derperforms No-Evolve on both benchmarks, de- creasing CL-Bench from29.54to27.96and SWE- bench Lite from63.67to61.67. This suggests that LLM-generated self-judgment can introduce misleading updates when it is not grounded in ex- ternal execution evidence. In contrast, Minimal and Standard feedback both use environment-provided signals and are more reliable overall. Standard feedback is generally preferable, but the useful granularity of feedback depends on the benchmark. On CL-Bench, Standard feed- back performs best, reaching34.02, while Mini- mal feedback reaches only29.86. This suggests that richer diagnostics are useful for adaptive rea- soning tasks, where understanding the reason for failure can help compile better harness guidance. On SWE-bench Lite, however, Minimal feedback reaches67.33, slightly above Standard feedback at67.00. One possible explanation is that detailed error messages can sometimes make the evolved harness overly tied to task-specific failures, while sparse pass/fail signals encourage more generaliz- able guidance. 5 Conclusion We formulated online harness learning as a setting for studying how frozen LLM agents can improve from sequential, one-shot executions. Through EVO-HARNESS, we instantiated this setting with context-to-harness skill compilation, which turns noisy execution contexts into a reusable external harness for future task solving. Beyond proposing a pipeline, our goal is to use the harness as an analyti- cal lens for understanding online self-improvement. This perspective moves self-evolving agent re- search beyond aggregate success rates toward a more principled study of how agents produce, or- ganize, transfer, and apply reusable experience in realistic task streams. Limitations Our study focuses on single-agent LLM systems in text-based, tool-use, web, command-line, and software-engineering environments. We do not evaluate embodied agents or multi-agent systems, where the harness may involve additional forms of interaction, coordination, and environmental feed- back. We also instantiate the harness with natural- language guidance; other harness formats, such as executable code skills or structured programs, are not covered in this study. References Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 techni- cal report. arXiv preprint arXiv:2303.08774. Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Alt- man, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925. Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Ar- nav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, and 1 others. 2025. Gepa: Reflec- tive prompt evolution can outperform reinforcement learning. arXiv preprint arXiv:2507.19457. Anthropic. 2025. Introducing claude opus 4.5. Anthropic. 2026a. Introducing claude opus 4.6. Anthropic. 2026b. Introducing claude opus 4.7. Yuanchen Bei, Tianxin Wei, Xuying Ning, Yanjun Zhao, Zhining Liu, Xiao Lin, Yada Zhu, Hendrik Hamann, Jingrui He, and Hanghang Tong. 2026. Mem-gallery: Benchmarking multimodal long-term conversational memory for mllm agents.arXiv preprint arXiv:2601.03515. Minghao Chen, Yihang Li, Yanting Yang, Shiyu Yu, Binbin Lin, and Xiaofei He. 2024. Automanual: Con- structing instruction manuals by llm agents via inter- active environmental learning. Advances in Neural Information Processing Systems, 37:589–631. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413. Shihan Dou, Ming Zhang, Zhangyue Yin, Chenhao Huang, Yujiong Shen, Junzhe Wang, Jiayi Chen, Yuchen Ni, Junjie Ye, Cheng Zhang, and 1 others. 2026. Cl-bench: A benchmark for context learning. arXiv preprint arXiv:2602.03587. Lutfi Eren Erdogan, Nicholas Lee, Sehoon Kim, Suhong Moon, Hiroki Furuta, Gopala Anumanchipalli, Kurt Keutzer, and Amir Gholami. 2025. Plan-and-act: Improving planning of agents for long-horizon tasks. arXiv preprint arXiv:2503.09572. Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang, Yuanchen Bei, Yankai Chen, Tao Feng, Xinyu Pan, Zhen Tan, Yu Wang, Tianxin Wei, and 1 others. 2026. Rethinking memory mechanisms of foun- dation agents in the second half. arXiv preprint arXiv:2602.06052. Guanyu Jiang, Zhaochen Su, Xiaoye Qu, and Yi R Fung. 2026. Xskill: Continual learning from experi- ence and skills in multimodal agents. arXiv preprint arXiv:2603.12056. Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R Narasimhan. 2024. SWE-bench: Can language mod- els resolve real-world github issues? In The Twelfth International Conference on Learning Representa- tions. Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. 2026. Meta- harness: End-to-end optimization of model harnesses. arXiv preprint arXiv:2603.28052. Minhua Lin, Hanqing Lu, Zhan Shi, Bing He, Rui Mao, Zhiwei Zhang, Zongyu Wu, Xianfeng Tang, Hui Liu, Zhenwei Dai, and 1 others. 2026a. Position: Agentic evolution is the path to evolving llms. arXiv preprint arXiv:2602.00359. Minhua Lin, Juncheng Wu, Zijun Wang, Zhan Shi, Yisi Sang, Bing He, Zewen Liu, Tianxin Wei, Zongyu Wu, Zhiwei Zhang, and 1 others. 2026b. Harness updating is not harness benefit: Disentangling evo- lution capabilities in self-evolving llm agents. arXiv preprint arXiv:2605.30621. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. Jiaqi Liu, Zipeng Ling, Shi Qiu, Yanqing Liu, Si- wei Han, Peng Xia, Haoqin Tu, Zeyu Zheng, Ci- hang Xie, Charles Fleming, and 1 others. 2026a. Omnimem: Autoresearch-guided discovery of life- long multimodal agent memory.arXiv preprint arXiv:2604.01007. Zewen Liu, Zhan Shi, Yisi Sang, Bing He, Minhua Lin, Tianxin Wei, Dakuo Wang, Benoit Dumoulin, Wei Jin, and Hanqing Lu. 2026b. Adaptive auto-harness: Sustained self-improvement for agentic system de- ployment on open-ended task streams. arXiv preprint arXiv:2606.01770. Xinghua Lou, Miguel Lázaro-Gredilla, Antoine Dedieu, Carter Wendelken, Wolfgang Lehrach, and Kevin P Murphy. 2026. Autoharness: improving llm agents by automatically synthesizing a code harness. arXiv preprint arXiv:2603.03329. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36:46534–46594. Mike A. Merrill, Alexander G. Shaw, Nicholas Car- lini, Boxuan Li, Harsh Raj, Ivan Bercovich, Lin Shi, Jeong Yeon Shin, Thomas Walshe, E. Kelly Buchanan, Junhong Shen, Guanghao Ye, Haowei Lin, Jason Poulos, Maoyu Wang, Marianna Nezhurina, Je- nia Jitsev, Di Lu, Orfeas Menis Mastromichalakis, and 66 others. 2026. Terminal-bench: Benchmark- ing agents on hard, realistic tasks in command line interfaces. Preprint, arXiv:2601.11868. Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Er- chao Zhao, Xiaoxi Jiang, and Guanjun Jiang. 2026. Trace2skill: Distill trajectory-local lessons into transferable agent skills.arXiv preprint arXiv:2603.25158. Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, Zihao Li, Yuanchen Bei, Jiaru Zou, Mengting Ai, Zhining Liu, Ting-Wei Li, and 1 others. 2026. Code as agent harness. arXiv preprint arXiv:2605.18747. Siru Ouyang, Jun Yan, Yanfei Chen, Rujun Han, Zifeng Wang, Bhavana Dalvi Mishra, Rui Meng, Chun- Liang Li, Yizhu Jiao, Kaiwen Zha, and 1 others. 2026. Skillos: Learning skill curation for self-evolving agents. arXiv preprint arXiv:2605.06614. Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T Le, Samira Daruki, Xiangru Tang, and 1 others. 2025. Reasoningbank: Scaling agent self-evolving with reasoning memory. arXiv preprint arXiv:2509.25140. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, and 1 others. 2024. Toolllm: Facilitating large language models to master 16000+ real-world apis. In International Conference on Learning Rep- resentations, volume 2024, pages 9695–9717. Libin Qiu, Zhirong Gao, Junfu Chen, Yuhang Ye, Weizhi Huang, Xiaobo Xue, Wenkai Qiu, and Shuo Tang. 2026. Autorefine: From trajectories to reusable expertise for continual llm agent refinement. arXiv preprint arXiv:2601.22758. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36:8634–8652. Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, and Jiecao Chen. 2025. Scaling long-horizon llm agent via context-folding. arXiv preprint arXiv:2510.11967. Mirac Suzgun, Mert Yuksekgonul, Federico Bianchi, Dan Jurafsky, and James Zou. 2025. Dynamic cheat- sheet: Test-time learning with adaptive memory. arXiv preprint arXiv:2504.07952. Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Mil- lican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805. Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, and 1 others. 2026. Kimi k2. 5: Visual agentic intelligence. arXiv preprint arXiv:2602.02276. Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, and 1 others. 2026. Skillx: Automatically constructing skill knowledge bases for agents. arXiv preprint arXiv:2604.04804. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Man- dlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. 2024. Agent workflow memory. arXiv preprint arXiv:2409.07429. Tianxin Wei, Noveen Sachdeva, Benjamin Coleman, Zhankui He, Yuanchen Bei, Xuying Ning, Mengting Ai, Yunzhe Li, Jingrui He, Ed H Chi, and 1 oth- ers. 2025. Evo-memory: Benchmarking llm agent test-time learning with self-evolving memory. arXiv preprint arXiv:2511.20857. Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, and 1 others. 2026. Skillrl: Evolving agents via recursive skill- augmented reinforcement learning. arXiv preprint arXiv:2602.08234. Wujiang Xu, Kai Mei, Hang Gao, Juntao Tan, Zu- jie Liang, and Yongfeng Zhang. 2025. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110. Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yux- uan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, and 1 others. 2026. Autoskill: Experience- driven lifelong learning via skill self-evolution. arXiv preprint arXiv:2603.01145. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024.τ-bench: A benchmark for tool- agent-user interaction in real-world domains, 2024. URL https://arxiv. org/abs/2406.12045, 1(3). Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei- Ying Ma, Jingjing Liu, Mingxuan Wang, and 1 others. 2025. Memagent: Reshaping long-context llm with multi-conv rl-based memory agent. arXiv preprint arXiv:2507.02259. Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. 2026a. Memskill: Learning and evolving mem- ory skills for self-evolving agents. arXiv preprint arXiv:2602.02474. Kangning Zhang, Shuai Shao, Qingyao Li, Jianghao Lin, Lingyue Fu, Shijian Wang, Wenxiang Jiao, Yuan Lu, Weiwen Liu, Weinan Zhang, and 1 others. 2026b. Mmskills: Towards multimodal skills for general visual agents. arXiv preprint arXiv:2605.13527. Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, and 1 others. 2025. Agentic context engineering: Evolving contexts for self-improving language models. arXiv preprint arXiv:2510.04618. Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. Expel: Llm agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632–19642. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Pro- ceedings of the AAAI conference on artificial intelli- gence, volume 38, pages 19724–19731. Shuyan Zhou. 2026. Webarena-infinity: Generating browser environments with verifiable tasks at scale. shuyanzhou.com. Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, and 1 others. 2024. Webarena: A realistic web environment for build- ing autonomous agents. In International Conference on Learning Representations, volume 2024, pages 15585–15606. Appendix A Benchmark Details Table 5 summarizes the benchmark settings. We use the same online protocol across benchmarks: tasks are processed as a stream, the solver uses the current harness, and the evolver updates the harness after observing task outcomes. Table 5: Benchmark statistics and evaluation signals. Benchmark#Tasks DomainEvaluation signal WebArena Inf.80 Web navigationState verifier Terminal-Bench89 CLI / scriptingDocker verifier SWE-bench Lite300 Software eng.Unit tests CL-Bench1899 Adaptive reasoning Rubric judge TAU-Bench165 Tool useState verifier WebArena Infinity. We use 80 hard tasks from the email-centered “superhuman-general” setting and related web applications. Tasks require stateful browser interaction, such as creating labels, chang- ing settings, using email filters, or coordinating cal- endar/email information. A programmatic verifier checks the final application state. This benchmark is useful for studying whether skills capture UI procedures. Terminal-Bench v2.We evaluate 89 tasks across software engineering, data processing, scientific computing, security, system administration, and related command-line domains. The solver works in an isolated shell environment and submits final artifacts or answers inside a task-specific Docker image. Final success is determined by the bench- mark’s executable verifier, typically a hidden or mounted test script run inside the container. SWE-bench Lite.We use 300 real GitHub issues from 12 Python repositories. The solver edits code and the final patch is evaluated by repository tests. CL-Bench. CL-Bench contains 1899 adaptive reasoning tasks across Domain Knowledge Reason- ing, Empirical Discovery & Simulation, Procedu- ral Task Execution, and Rule System Application. Each task is graded with rubric criteria. TAU-Bench.TAU-Bench contains 165 customer- service tool-use tasks across airline and retail do- mains. Evaluation checks the final database state. B Harness Artifacts Examples This section shows representative learned artifacts without enumerating the learned inventory. For task-level comparison, we align the same task iden- tifiers between the no-evolve and evolved runs. We only call a case a direct improvement when the baseline run fails and the evolved run succeeds. This is not a single-skill causal ablation: several skills may be injected together, and some bench- marks do not record injected skills in every result file. The examples therefore use a structured fail- ure/success format: task, verifier signal, trajectory summary, learned skill content, and a short take- away. WebArena: verified setting •Task:task_h10; create “Legal” and disable “Support Ticket”. •Baseline: verifier found “Support Ticket” still enabled, despite the trajectory claiming both subgoals were com- plete. •Evolved: verifier passed; the trajectory returned to the Auto Labels list and checked the existing label state. •Learned skill: toggle an existing auto label off; do not delete it; verify creation and disabling independently. Core point. The harness learned a concrete UI distinction: create-new-label and disable- existing-label are separate operations. WebArena: grounded recipients •Task:task_h65; reply to Kevin Zhao with CTO and VP Engineering on C. • Baseline: sent to plausible names but wrong domains (acme.com instead of acmecorp.com). •Evolved: verifier passed with the required role-derived recipients in C. •Learned skill: prefer authoritative role sources and preserve exact email addresses across apps. Core point.The harness learned identity grounding, not merely email composition. Terminal-Bench: executable recovery •Task:adaptive-rejection-sampler; pass the Docker verifier for an R sampler. •Baseline: formal tests failed after repeated reruns of the same R test command. •Evolved: passed after diagnosing timeout, missing dependency, formal-test, and sorted-vector failures. •Learned skill: after repeated similar failures, stop, shrink the case, build a minimal reproducer, then add complexity back. Core point. The harness learned a recovery policy for executable feedback loops. SWE-bench: test-grounded repair •Task:django__django-11133; fix Django response handling without regressing existing behavior. •Baseline: timed out after source search and failed test attempts. •Evolved:passedFAIL_TO_PASS: 1/1and PASS_TO_PASS: 64/64. •Learned skill: read complete failing-test assertions, match exact expected behavior, then run targeted tests before submission. Core point. The harness learned to anchor patches in executable test expectations. TAU-Bench: complete tool execution •Task:retail_task_5; exchange multiple items from one order. •Baseline: database verifier failed, consistent with par- tial or wrong-tool execution. •Evolved: verifier passed after retrieving the order and applying all requested item changes. • Learned skill: enumerate every requested operation, choose the tool by order status, and verify the response covers the full list. Core point. The harness learned complete tool- call execution. C Which Tasks Are Easier to Improve? The most informative pattern is not the total number of fail-to-pass cases, but the kind of task that turns around after harness evolution. Table 6 groups matched fail-to-pass cases by their observable fail- ure mode and gives representative task identifiers from the runs in outputs. D What the Harness Actually Learns Figure 7 shows real skill excerpts selected with matched fail-to-pass support. The evidence line reports the relevant turnaround family in the runs, while the rule line gives the core learned behavior. E Prompt Templates The prompts follow the same context-to-harness compilation logic across experiments: the solver re- flects on an execution context to propose candidate lessons, and the evolver curates these candidates against the existing harness. Below, we provide representative prompt templates that capture this common proposal-and-curation process. E.1 Solver-Side Proposal Prompt Proposal prompt, compressed Inputs: evaluation result, verifier details or rubric feedback, trajectory signals, compressed trajectory, and related existing skills. Analyze the execution. For each distinct issue, iden- tify what went wrong and the concrete missing action, command, tool call, navigation step, or domain rule. Propose a reusable skill. Choose a broad topic, decideNEW,ENHANCE, orNONE, write a trigger-style description, and provide short bullet-point content with techniques and gotchas. Filter aggressively. Skip generic advice, basic tool usage, exact task replay, and skills that would not help unseen tasks. E.2 Evolver-Side Curator Prompt Curator prompt, compressed Inputs: topic, current skill library, budget, and pro- posals from the latest batch. For each proposal, chooseACCEPT,MERGE, orSKIP. Prefer merging over duplication, respect the budget, require a clear trigger description, keep content short and actionable, and apply a generalizability test such as usefulness for multiple unseen tasks. E.3 General Skill Curator General skill curator, compressed Analyze repeated patterns across tasks rather than within a single topic. Create or update a general skill only when a pattern appears across multiple contexts. General skills must avoid context-specific references and encode procedures that can guide planning, veri- fication, recovery, or tool use across tasks. F Experimental Details All experiments use the same open-source pipeline with benchmark-specific adapters. WebArena uses browser workers with persistent application state. Terminal-Bench uses task-specific Docker images and executable verifiers. SWE-bench uses isolated repository environments and unit-test evaluation. TAU-Bench uses the official tool-use environment and verifier. CL-Bench uses a rubric-based judge pipeline. The harness is stored as Markdown skill files with lightweight YAML metadata, making the learned artifact inspectable and easy to transfer across runs. Each entry contains a trigger describ- ing when it should be retrieved, a short actionable rule or procedure, optional evidence linking it to prior executions, and a scope indicating whether it is intended as cross-task guidance or task-type Table 6: Improvement hotspots from matched baseline-fail/evolved-pass cases. BenchmarkEasier-to-improve tasks Representative casesWhy skills help WebArenaStateful UI tasks with ex- act recipients, labels, set- tings, and mail state. h10: disable an existing auto label; h65: preserve corporate email do- mains;h73: archive all qualifying read-important emails. The verifier checks final app state, so procedural memory about where to click must be coupled with exact state confir- mation. Terminal- Bench Tasks with executable feedback loops and recov- erable failures. adaptive-rejection-sampler, dna-insert,build-pmars, regex-chess. The solver can observe Docker- test failures; process skills help it pivot, reduce scope, and verify artifacts before submission. SWE-bench Lite Mature repositories with stable conventions, espe- cially Django and Sympy. django__django-11133, django__django-17087, sympy__sympy-19254. The issue often hides exact ex- pectations in tests or framework internals; topic skills preserve these conventions across later tasks. TAU-BenchRetail exchanges/returns and airline reservation ed- its with multiple con- straints. retail_task_5 , retail_task_91, airline_task_13. The state verifier rewards com- plete tool-call execution. Skills prevent partial updates, wrong mutation tools, and guessed item or flight identifiers. CL-BenchRule-application,pro- cedural,and domain- knowledge contexts with local rubrics. Context categories with many turnarounds include Rule System Application, Procedural Task Ex- ecution, and Domain Knowledge Reasoning. Context skills behave like small local manuals: exact extraction rules, role boundaries, output schemas, and calculation con- ventions. guidance. For harness selection, we use Claude Sonnet 4.5 across all experiments to retrieve rel- evant skills from the current harness before task execution. Unless otherwise specified, we use a batch size of 16. We set the maximum number of general skills and the maximum number of skills under each task-type topic to 5. This budget encourages the evolver to merge overlapping guidance and avoid accumulating overly detailed or task-specific records. G Potential Risks EVO-HARNESS evolves skill libraries that are in- jected into agent prompts as procedural guidance. The learned skills are designed to capture task- solving patterns such as navigation flows, tool- use procedures, verification steps, and recovery strategies. In our experiments, all benchmarks are run in sandboxed environments with no access to real user data, external accounts, or production systems. Therefore, the evaluated setting poses minimal practical risk beyond standard benchmark execution. H Use Or Create Scientific Artifacts Our work uses five established public benchmarks: WebArena-Infinity, TerminalBench2, SWE-bench Lite, CL-Bench, and TAU-Bench. We do not mod- ify the benchmarks themselves. We use founda- tion models (Claude and other models) via their standard AWS Bedrock APIs. We will release the EVO-HARNESS pipeline code and all evolved skill libraries upon publication. H.1 Cite Creators Of Artifacts All benchmarks and models are properly cited in the main text. H.2 Discuss The License For Artifacts All benchmark resources used in this work are publicly available under their respective research or open-source licenses. In particular, WebArena Infinity and TAU-Bench are released under MIT licenses, WebArena is released under Apache- 2.0, and CL-Bench is released under a custom evaluation-only license. Our pipeline code will be released under the MIT License. WebArena: UI-state verification Support. Settings-state tasks account for 24% of We- bArena fail-to-pass cases; related UI skills appear in 72% of improved runs. Guidance. Treat creation and disabling as separate opera- tions; after editing an auto label, return to the settings list and verify its final state. WebArena: grounded identity use Support. Identity-grounding failures account for 32% of WebArena fail-to-pass cases. Guidance. Use authoritative sources for role contacts and preserve exact email addresses; do not infer or normalize domains. Terminal-Bench: executable recovery Support. Executable-feedback and recovery patterns ac- count for 100% of Terminal-Bench fail-to-pass cases. Guidance. Stop rerunning the same command, isolate a minimal failing case, fix the core issue, and then rebuild toward the full solution. SWE-bench: test-grounded repair Support. Django and Sympy account for 66.7% of SWE- bench fail-to-pass cases. Guidance. Read the full failing-test assertions before editing; match exact expected strings, error identifiers, and output structures. TAU-Bench: complete tool execution Support. Retail workflows account for 85.2% of TAU- Bench fail-to-pass cases; this guidance appears in 100% of improved runs. Guidance. Enumerate every requested operation, choose tools according to the current state, and verify that the response covers the full request. CL-Bench: constrained extraction Support. Rule, procedural, and domain-knowledge con- texts dominate CL-Bench fail-to-pass cases. Guidance. For extraction tasks, return the requested source text verbatim and match only the exact term or field specified by the prompt. Figure 7: Representative learned harness excerpts across benchmarks. Each card pairs a recurring fail-to-pass pattern with the corresponding evolved guidance. Percentages summarize matched baseline-fail/evolved-pass cases and contextualize the examples rather than establish single-skill causality. H.3 Artifact Use Consistent With Intended Use All benchmarks are used for their intended purpose of evaluating agent capabilities. The foundation models are used via their standard APIs in accor- dance with their terms of service. H.4Data Contains Personally Identifying Info Or Offensive Content The benchmarks use synthetic or sanitized scenar- ios. SWE-bench uses public GitHub issues. To our knowledge, none contain personally identifying in- formation or offensive content. I Computational Experiments I.1 Compute Budget The full experimental suite is computationally ex- pensive because it covers multiple benchmarks, solver models, evolution settings, and ablation con- figurations. Across all main experiments and di- agnostic runs, the total API cost was on the or- der of$100K. For a single benchmark configu- ration, we typically run 16 parallel workers. De- pending on benchmark complexity, task timeout, and environment overhead, one full run takes ap- proximately 2–8 hours of wall-clock time. Web and tool-use benchmarks are usually faster, while software-engineering and command-line tasks tend to take longer because they require repository setup, Docker execution, test running, or verifier calls. I.2 Experimental Setup And Hyper-params Key hyperparameters are described in Appendix F. We use fixed random seeds (42) for task shuffling to ensure reproducibility. I.3 Descriptive Statistics We report pass rate (%) as the primary metric across all benchmarks. For statistical reliability, we report results over the full task set for each benchmark rather than subsampling. The performance is aver- aged over three runs. I.4 Parameters For Packages The project primarily used AWS Bedrock API for foundation model inference, together with standard AWS infrastructure services such as EC2, S3, EB- S/FSx, CloudWatch, IAM, and VPC. The software environment used Python 3.10+, boto3 1.34+, boto- core 1.34+, PyTorch 2.0+, CUDA 12.0+, Docker, and Ubuntu 24.04. J AI Assistants In Research Or Writing This paper studies LLMs as the research object and uses frozen LLM backbones for experimental inference. These models are not used to create new benchmark labels or to fabricate experimental results. Reported numerical results come from the experimental code and are checked by the authors. AI assistant tools were used for polishing manuscript. They were not used to decide sci- entific claims, select reported results, or replace author verification.