Source-linked AI summary
MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
Huawei Lin, Peng Li, Jie Song, Fuxin Jiang, Tieying Zhang
TL;DR
Existing skill systems lack structured memory, validation, and lifecycle management, limiting reusable improvement across tasks. MUSE-Autoskill unifies skill creation, memory, management, evaluation, and refinement, leading compared agents on reported benchmarks with 53.42% all-task self-created-skill accuracy and 51.90% transfer into Hermes.
Problem
Existing approaches leave skill creation disconnected from execution, per-skill memory absent, skills static and unvalidated, and long-horizon context poorly managed.
Method
MUSE-Autoskill unifies runtime skill creation, multi-level memory, unit-test evaluation, automatic refinement, and adaptive context management across tasks.
Results
MUSE-Autoskill leads compared self-creating agents on reported benchmarks, including a 53.42% all-task self-created-skill result and 51.90% transfer into Hermes.
Takeaways & Limitations
The results support skill packages as a practical unit for accumulating and reusing agent experience across repeated tasks.
Takeaways & Limitations
The evaluation excludes 19 SkillsBench tasks, covers incomplete skills, evaluates transfer only into Hermes, and uses few runs with wide task-level confidence intervals.
Abstract
from arXiv · showhide
Large language model (LLM) agents rely on reusable skills to solve complex tasks, but existing skill creation approaches often treat skills as isolated, static artifacts, limiting reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that creates, reuses, and refines skills under a unified lifecycle: creation, memory, management, evaluation, and refinement. MUSE creates skills on demand, stores them across tasks, retrieves them through a skill catalog, and accumulates per-skill experience for later reuse and adaptation. Across the main reported settings on SkillsBench and SkillLearnBench, MUSE-Autoskill outperforms Hermes, Codex, and Claude Code. On SkillsBench, its self-created skills surpass human-authored skills on the successfully covered subset (85.24% vs. 81.17%), showing that lifecycle-managed skills can distill agent experience into highly effective reusable assets; MUSE-created skills also transfer to Hermes more effectively than Codex- or Claude-created skills, reaching 51.90% accuracy under transfer. These results highlight the importance of treating skills as long-lived, experience-aware, and testable assets.
1 Introduction
The paper reframes skills as long-lived, experience-aware assets managed through a unified lifecycle of creation, memory, management, evaluation, and refinement. MUSE-Autoskill implements this lifecycle and achieves strong results across SkillsBench and SkillLearnBench, including effective cross-agent skill transfer.
- Limits of AutoSkill: Existing AutoSkill methods leave four gaps: creation–usage mismatch, absent per-skill memory, static unvalidated skills, and poor long-context handling.These gaps concern runtime context, accumulated experience across tasks, unit-test-driven evaluation and refinement, and conversation-history limits.
- Skill lifecycle: MUSE-Autoskill unifies skill creation, memory, management, evaluation, and refinement into a lifecycle for long-lived, evolving agent assets.Skills are created on demand, stored with experience and metadata, retrieved contextually, validated through tests and runtime feedback, and refined as evidence accumulates.
- MUSE-Autoskill framework: MUSE-Autoskill integrates runtime skill creation, multi-level memory, adaptive context compression, unit-test evaluation, sandbox feedback, and cross-agent transfer.Its uniquely skill-level memory accumulates per-skill experience across tasks and informs future invocations.
- Results: 59.67% human-skill accuracy on SkillsBench’s 75-task common set was MUSE-Autoskill’s best result, a +12.72 pp lift over its no-skill accuracy.MUSE-Autoskill also achieved 46.95% no-skill accuracy and the strongest self-created-skill result under all-task scoring at 53.42%.
- Results: 51.90% Hermes accuracy with MUSE-Autoskill-created skills exceeded Hermes with human skills at 48.02%, while SkillLearnBench results reached 72.0% with human skills and 48.0% with self-created skills.MUSE-created skills also had the lowest run-to-run reward dispersion among compared self-created-skill conditions.
2 Related Work
Prior work on automatic skill systems can be organized by lifecycle coverage and training requirements, spanning training-free, reinforcement-learning-based, and benchmark-driven approaches. MUSE-Autoskill distinguishes itself by unifying all five lifecycle stages in one training-free framework with persistent skill memory, test-triggered refinement, and empirically validated cross-agent transfer.
- Organizing prior work: Automatic skill systems are compared by coverage of creation, memory, management, evaluation, and refinement, and by whether they require training beyond inference.Table 1 defines memory as persistent per-skill experience across tasks and distinguishes training-free methods from those using fine-tuning or reinforcement learning.
- Training-free skill systems: Voyager exemplifies training-free skill systems by maintaining executable-code skills with self-verification and iterative refinement from environment feedback.Follow-up work extends this paradigm toward general-purpose agents, including AutoSkill.
- Reinforcement-learning skill systems: Reinforcement-learning approaches jointly optimize skill behavior, policy action, skill-bank editing, selection, utilization, or distillation under downstream or task-outcome rewards.SkillMaster, Skill1, and SkillOS represent this concurrent direction with different allocations of acting, editing, execution, and curation responsibilities.
- Benchmarks: Recent benchmarks stress different lifecycle stages: SkillsBench measures end-to-end task accuracy, SkillRet evaluates large-scale retrieval, and SkillLearnBench and LifelongAgentBench examine continual or lifelong acquisition.SkillsBench uses diverse Docker-evaluated real-world tasks and is adopted for the paper’s experiments.
- MUSE-Autoskill: MUSE-Autoskill brings all five lifecycle stages together in a single training-free framework, adding per-skill experience memory, unit-test-triggered refinement, and empirically validated cross-agent transfer without modification.The framework differs from methods addressing creation or refinement in isolation and is described as the only general-purpose method to empirically validate cross-agent skill transfer.
3 MUSE-Autoskill Agent
MUSE-Autoskill is a skill-centric agent framework that dynamically creates, reuses, evaluates, and refines skills within a unified lifecycle. Its iterative Planning–Action–Observation loop uses memory, skill packaging, and validation to support adaptive execution and reuse.
- Framework overview: MUSE integrates skill creation, execution, memory, management, and evaluation into a unified agent loop for dynamically creating, reusing, and refining skills.The framework organizes skills into five lifecycle stages: creation, memory, management, evaluation, and refinement.
- Agent loop: The agent repeatedly plans, invokes retrieved or built-in skills, and observes execution results to refine behavior, handle errors, and complete complex multi-step tasks.Planning uses problem decomposition and memory; Action invokes skills; Observation feeds execution results into subsequent decisions.
- Skill representation: A skill is a structured directory centered on SKILL.md, with optional scripts/, resources/, and tests/ components executed through a unified interface.SKILL.md specifies the skill’s name, description, inputs, and outputs, while the runtime selectively reads resources or runs scripts.
- Skill memory and reuse: Reused skills reduce repeated exploration by letting the agent load an interface first and access full procedures or scripts only when needed.Skill-level memory stores skills, metadata, descriptions, inputs, and usage history to support retrieval for new tasks.
- Skill creation: When existing skills are insufficient, skill_create turns a high-level specification into a complete executable package by generating its interface, internal structure, files, and optional tests.The generated package can include scripts/, resources/, and tests/ for code-backed skills.
4 Experiments
Across SkillsBench and SkillLearnBench, MUSE-Autoskill achieves the strongest reported benchmark results under no-skill, human-skill, and self-created-skill conditions. Its generated skills also improve transfer and covered-task efficiency, while remaining limited by uncovered tasks and occasional regressions.
- Evaluation Protocol: Evaluation uses 375 SkillsBench runs across 75 tasks and five repetitions, while SkillLearnBench evaluates 100 verified instances once; operational failures count as zero.Self-created-skill settings use the same denominator as corresponding no-skill and human-skill settings.
- Main Results: MUSE-Autoskill achieves the highest no-skill, human-skill, and self-created-skill accuracy in both benchmarks, reaching 59.67% with human skills on SkillsBench and 72.0% on SkillLearnBench.Human skills improve every agent on both benchmarks.
- Self-Created Skills: 85.24% covered-task accuracy exceeds MUSE-Autoskill’s 81.17% human-skill accuracy, while all-task scores remain lower because 28–31 tasks lack usable generated skills.Codex and Claude Code reach 75.83% and 75.45% on covered tasks.
- Skill Transfer: 51.90% transfer accuracy raises Hermes from 37.24% without skills and exceeds Hermes with human skills at 48.02%.The corresponding improvement is +14.66 pp, while MUSE-Autoskill and Codex have the same 47-task source coverage in this run.
- Cost and Efficiency: 85.24% accuracy on MUSE-Autoskill’s 47 covered tasks accompanies 434.7 s median latency and 499K median tokens, versus 74.92% without skills and 81.17% with human skills.Skill creation costs a median 363.6K tokens and 156.3 s per covered task; reuse breaks even after roughly three reuses for tokens and one for latency.
- Limitations: 28 of 75 tasks remain uncovered, concentrated in Ops & Planning (12) and Data Analysis (9), and one HVAC-control regression falls from 80% to 20%.The regression stems from calibration variance producing gains outside the verifier’s stability margin when the skill is reapplied.
5 Real-World Deployment and Impact
MUSE-Autoskill’s skill-centric lifecycle is being adopted in production systems as a shared capability unit across agents and users. These deployments extend the framework beyond benchmarks to hosted skill storage, evaluation, governance, memory, management, and refinement.
- Production deployments: SkillMarket turns successful trajectories into reusable, self-tested skill packages for end users without manual authoring.Planned releases add skill management and updating, enabling deployed skills to be versioned and refined.
- Production deployments: SkillHub provides teams a hosted service covering skill creation, evaluation, memory, management, and refinement, with shared storage and governance.It preserves accumulated per-skill experience alongside the skills themselves.
- Future impact: Developers may increasingly compose and version skills as workflow primitives whose bundled tests and memory make workflows self-documenting and easier to maintain.This shifts maintenance from bespoke glue code toward a shared, continuously evaluated skill ecosystem.
6 Conclusion
MUSE-Autoskill presents a unified lifecycle for acquiring, reusing, evaluating, executing, remembering, managing, and refining structured skills. It reports leading results on SkillsBench and SkillLearnBench, supporting skill packages as practical units for accumulating and reusing agent experience.
- Framework: MUSE-Autoskill improves task-solving by acquiring, reusing, and refining skills through a unified lifecycle of structured, reusable procedures.This lifecycle helps avoid rediscovering commands, file formats, and validation steps on later runs.
- Framework: MUSE-Autoskill integrates skill creation, evaluation, execution, memory, and management around minimal built-in skills such as skill_- create and web_search.The framework is designed to support reusable skills across repeated tasks.
- Results: 59.67% is MUSE-Autoskill’s strongest human-skill accuracy on SkillsBench, while 53.42% is its strongest all-task self-created-skill result.These are reported as the strongest results for the respective SkillsBench settings.
- Results: 51.90% is MUSE-Autoskill’s best transfer into Hermes, and it leads compared self-creating agents on 100 verified SkillLearnBench instances.Together with production deployments, these results support skill packages as practical units for accumulating and reusing agent experience.
Limitations
The evaluation is limited by partial benchmark coverage, incomplete self-created skills, single-trajectory and same-task testing, and transfer assessed only into Hermes. Generated skills may hallucinate unsupported or brittle assumptions, so stronger provenance, adversarial testing, and human review remain necessary.
- Evaluation scope: 75 of 94 SkillsBench tasks were evaluated because all four runtimes could run only 75 locally, while excluded tasks often have more complex Docker environments.MUSE-Autoskill and Codex produced usable skills for 47 of 75 tasks, compared with 44 for Claude Code.
- Evaluation scope: 47 of 75 tasks had usable skills from MUSE-Autoskill and Codex, whereas Claude Code covered 44; each skill came from a single source trajectory and transfer was tested only into Hermes.The evaluation also used 5 runs per task, limiting the breadth of evidence.
- Evaluation protocol: One successful Phase 1 trajectory and re-evaluation on the same task may overstate within-task gains, despite deterministic verification and no task-specific ground truth being fed into the skill.SkillLearnBench uses a separate 100-instance, one-run-per-instance protocol and should be treated as corroborating rather than pooled evidence.
- Reliability risks: Generated skills or invoking agents may hallucinate unsupported facts, brittle file-path or API assumptions, and heuristics overfit to a source trajectory.Unit tests, sandbox execution, verifier feedback, and leakage checks reduce but do not eliminate this risk.
- Reliability risks: High-impact deployments require stronger provenance tracking, adversarial tests, and human review before skills are shared across teams or exposed to user-facing workflows.These safeguards address residual hallucination and overfitting risks rather than eliminating them.
A Selected Task List
The four-agent comparison uses a 75-task common set from SkillsBench, organized into four super-domains. Tasks retain their original categories when available, with primary task tags used as fallback.
- Task set: 75 SkillsBench tasks form the common set used for the four-agent comparison in Figure 1.Table 9 identifies this as the selected task list for the comparison.
- Super-domains: 18 tasks belong to Science & Engineering, 18 to Data Analysis, 14 to Document Processing, and 25 to Ops & Planning.These four super-domains comprise all 75 selected tasks.
- Categorization: Tasks use the original SkillsBench category when available and the primary task tag when the category field is absent.This procedure determines the category labels used for grouping.
- Task examples: Examples include paper-anonymizer and pdf-excel-diff in Document Processing, parallel-tfidf-search in Data Analysis, and pedestrian-traffic-counting in Science & Engineering.The listed task rows show task IDs, SkillsBench categories, and assigned super-domains.
B FDABench Supplementary Evaluation … E Hyperparameters and Runtime Configuration
The supplementary sections report FDABench performance, define the skill package and filesystem conventions, and document fixed runtime settings used across SkillsBench experiments.
- B FDABench Supplementary Evaluation: FDABench-Full covers 579 single-choice, 760 multiple-choice, and 668 report task instances across three heterogeneous data-analysis families.The families use exact-match grading for choice tasks and database-backed evidence for report generation.
- B FDABench Supplementary Evaluation: 74.10% EX on single-choice tasks was achieved with 281.4s average latency and 179.6M aggregate tokens (310.2K per task).FDABench is included as a supplementary stress test rather than a headline comparison.
- C Skill Package Schema: A skill is a kebab-case directory containing a top-level SKILL.md with YAML frontmatter, with optional scripts/, tests/, resources/, and references/ subdirectories.Skills without code may consist of SKILL.md alone, which is the dominant practical pattern.
- C Skill Package Schema: The runtime eagerly exposes only each skill’s name and frontmatter description, loading the SKILL.md body through read_skill after catalog routing selects it.This two-stage lookup keeps per-call input cost flat as the skill bank grows.
- C Skill Package Schema: Each skill has a lazily created sibling .memory.md file that stores notes, lessons, and usage observations outside the published skill surface.Keeping experience outside the transferable skill prevents memory from being transferred with the skill artifact.
- D File-System Layout: The default persistent agent state resides under $HOME/.autoskill, while each task receives a UUID-like session directory containing prompts, inputs, outputs, logs, events, memory, and serialized state.The runtime also uses isolated /sandbox inputs and outputs for each invocation.
- E Hyperparameters and Runtime Configuration: All runtime constants were fixed across the 75-task common set without per-task tuning.The constants are organized into compression, timeout, retry-and-verification, and backbone-and-agent groups.
- E Hyperparameters and Runtime Configuration: All four agents share gpt-5.5-2026-04-24, and every task runs 5 times in independent Docker containers under the SkillsBench harness’s wall-clock budget.Accuracy differences therefore reflect agent-system choices such as prompts, tools, compression, context handling, and skill loading rather than model-backbone differences.
F SkillLearnBench Memory-System On/Off Ablation … I Per-Domain Accuracy with Standard Deviation
Across four analyses, MUSE-Autoskill benefits from memory, controlled context compression, and reusable skills, while domain results and token diagnostics characterize where these gains arise and how variable performance remains.
- F SkillLearnBench Memory-System On/Off Ablation: +14.0 percentage points raises round 2 verifier-feedback accuracy from 60/100 to 74/100 with the memory system enabled.Round 1 accuracy also rises from 48/100 (48.0%) to 64/100 (64.0%).
- F SkillLearnBench Memory-System On/Off Ablation: Round 1 execution cost falls from 524.4k to 467.3k tokens/run, 12.7 to 11.4 turns/run, and 419.1s to 347.7s with memory enabled.In round 2, tokens increase from 408.4k to 423.9k while turns remain nearly unchanged at 11.6 to 11.7.
- G Compression Algorithm: Context compression runs at every ReAct turn but triggers summarization only after the token estimate crosses COMPRESS_TOKEN_THRESHOLD.Level 1 summarizes oversized middle nodes; Level 2 collapses the middle span into one summary node.
- G Compression Algorithm: Summarization preserves revisited early-context facts at ∼1/10 the token cost, whereas truncation can force wasteful rediscovery on multi-step tasks.Agents revisit early-context facts roughly 30–40% of the time in preliminary experiments.
- H Detailed Token Breakdown: MUSE-Autoskill and Codex show substantial prompt-cache use, with median cached input larger than median fresh input in every reported condition.Human skills increase MUSE-Autoskill’s median token footprint mainly through additional skill-catalog and skill-body context.
- I Per-Domain Accuracy with Standard Deviation: Per-domain standard deviation is computed across task-level means, so high σ reflects variation among tasks rather than direct run-to-run noise.Each task mean is averaged over 5 runs.
- I Per-Domain Accuracy with Standard Deviation: Skills improve all four domains for all four agents in aggregate, while MUSE-Autoskill leads human-skill accuracy in three domains and Claude Code leads Data Analysis.MUSE-Autoskill leads Science & Engineering, Document Processing, and Ops & Planning; Hermes gains most in Science & Engineering, and the largest lift for several agents occurs in Ops & Planning.
J Latency and Turn-Count Distribution
Across 375 runs per agent-condition cell, Hermes has the fastest median latency, while MUSE-Autoskill uses deeper loops but avoids Codex’s long tail. Human skills reduce median wall-clock latency for every agent, although turn counts do not uniformly decrease.
- Latency distribution: Hermes is fastest by median latency, followed by Claude Code, MUSE-Autoskill, and Codex.Latency and turn-count percentiles are computed over 75 tasks × 5 runs, or 375 runs, per agent-condition cell.
- Latency distribution: MUSE-Autoskill runs deeper loops than Hermes and Claude Code, but its latency tail is shorter than Codex’s.Claude Code also has a longer high-percentile tail than Hermes despite a comparable median.
- Human-skill condition: 354.0 s to 327.3 s, 1013.6 s to 869.5 s, 347.3 s to 291.4 s, and 747.6 s to 730.8 s are the median-latency reductions for Hermes, Codex, Claude Code, and MUSE-Autoskill, respectively.Human skills reduce median latency for every agent while improving accuracy.
- Human-skill condition: Turn counts do not uniformly decrease, so the efficiency claim concerns wall-clock latency rather than fewer ReAct steps.The reported latency reductions therefore cannot be interpreted as uniform reductions in agent turn counts.
K Self-Created Skill Stability
MUSE-Autoskill’s self-created skills show run-to-run stability on each agent’s covered subset, measured across five independent runs. The paper attributes this pattern to executable procedural details distilled from successful trajectories.
- Stability measurement: Run-to-run stability is evaluated using mean reward, standard deviation, and mean absolute deviation across five independent runs per task, averaged over each covered subset.Lower standard deviation and lower mean absolute deviation indicate more consistent repeated-run rewards, regardless of the mean reward level.
- Stability measurement: Table 15 counts non-constant tasks and low-variance tasks, where low variance means per-task reward standard deviation at most 0.1.Non-constant tasks have five rewards that are not identical.
- Procedural explanation: MUSE-Autoskill-created skills often encode executable procedures rather than general descriptions, preserving concrete details from successful trajectories.Examples include command sequences, file paths, output schemas, validation checks, and task-specific failure modes, which narrow run-specific choices.
L Skill-Generation Failures: The 28 MUSE-Uncovered Tasks
MUSE-Autoskill produced no usable self-created skill for 28 of 75 tasks, exposing coverage limits concentrated in Ops & Planning and Data Analysis. These failures indicate that skill creation depends on Phase 1 finding reusable source trajectories.
- Uncovered tasks: 28 of 75 tasks produced no usable self-created skill under the strict protocol, contributing 0% to the all-task average.The uncovered tasks characterize current limits of inference-time skill synthesis.
- Uncovered tasks: The uncovered tasks concentrate in Ops & Planning and Data Analysis, with smaller clusters in Science & Engineering and Document Processing.This distribution supports the paper’s bottleneck analysis of self-created skill coverage.
- Bottleneck and direction: Coverage depends on Phase 1 exploration finding a reusable source trajectory, motivating extraction of partial skills from failed trajectories.Such partial skills could capture diagnostic moves that worked even when the overall run failed.