CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Hanrong Zhang; Shicheng Fan; Henry Peng Zou; Yankai Chen; Zhenting Wang; Jiayu Zhou; Chengze Li; Wei-Chieh Huang

cs.AIApr 2, 2026

CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Authors: Hanrong Zhang, Shicheng Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayu Zhou, Chengze Li, Wei-Chieh Huang,

Organizations: Steve · University of Illinois Chicago · MBZUAI · McGill University ·

Abstract

Anthropic proposes the concept of skills for LLM agents to tackle multi-step professional tasks that simple tool invocations cannot address. A tool is a single, self-contained function, whereas a skill is a structured bundle of interdependent multi-file artifacts. Currently, skill generation is not only label-intensive due to manual authoring, but also may suffer from human--machine cognitive misalignment, which can lead to degraded agent performance, as evidenced by evaluations on SkillsBench. Therefore, we aim to enable agents to autonomously generate skills. However, existing self-evolving methods designed for tools cannot be directly applied to skills due to their increased complexity. To address these issues, we propose CoEvoSkills, a self-evolving skills framework that enables agents to autonomously construct complex, multi-file skill packages. Specifically, CoEvoSkills couples a Skill Generator that iteratively refines skills with a Surrogate Verifier that co-evolves to provide informative and actionable feedback without access to ground-truth test content. On SkillsBench, CoEvoSkills outperforms five baselines on both Claude Code and Codex, and generalizes strongly to six additional LLMs. The code is publicly available at https://github.com/Zhang-Henry/CoEvoSkills.

Figures & tables

Figure 1: Tool–skill difference illustration.

Figure 2: Skill quality improvement across 5 evolution rounds. CoEvoSkills surpasses human-curated skills within 5 evolution iterations.

Figure 3: Overview of the CoEvoSkills co-evolutionary framework. The Skill Generator and Surrogate Verifier co-evolve through iterative refinement. The verifier provides structured failure feedback to drive skill improvement, while a ground-truth oracle test returns only an opaque pass/fail signal, triggering test escalation and ensuring strict information isolation.

Figure 4: Skill quality comparisons with baselines on SkillsBench (Claude Opus 4.6 + Claude-Code). Error bars:

\pm

1 std over 5 runs.

Component	Setting
Backbones	Claude Opus 4.6, GPT-5.2
Surrogate verifier model	Same as evolution backbone (Claude Opus 4.6 / GPT-5.2)
Ground-truth oracle test agent	Claude-Code (Claude Opus 4.6) / Codex (GPT-5.2)
Evolution stage	$K{=}5$ Ground Truth Oracle rounds, $M{=}15$ surrogate repair retries
Evolution runtime	timeout multiplier $5\times$ (effective 3000s/task), 4 parallel workers
Evaluation stage	timeout 7200s/task, 10 parallel workers

Table 1: Shared configuration for evolution and evaluation.

Model	Agent harness
Claude Opus 4.6	Claude-Code ( Anthropic, 2025a )
GPT-5.2	Codex ( OpenAI, 2025a )
Claude Sonnet 4.5	Claude-Code ( Anthropic, 2025a )
Claude Haiku 4.5	Terminus-2 ( Merrill et al., 2026 )
Qwen3 Coder	Terminus-2 ( Merrill et al., 2026 )
DeepSeek V3	Terminus-2 ( Merrill et al., 2026 )

Table 2: Agent harness used for each model in ground-truth oracle evaluation.

Figure 5: Cross-model skill transferability on SkillsBench. Skills evolved by Claude Opus 4.6 are transferred to six additional models spanning five providers. Each pair of bars shows the no-skill baseline (red) and the with-skills pass rate (blue). Delta annotations indicate absolute improvement. All models benefit substantially, with gains of approximately 35–44 percentage points across models, confirming that the evolved skills encode reusable task structure rather than model-specific artifacts.

Self-Evolved Skills
Model	With skills	No skill	$\Delta$
Claude Opus 4.6 (self-evolved)	71.1	30.6	+40.5
GPT-5.2 (self-evolved)	69.8	29.6	+40.2
Cross-Model Transfer (Opus 4.6 Evolved Skills)
GPT-5.2	65.0	29.6	+35.4
Claude Sonnet 4.5	63.1	20.0	+43.1

Table 3: Cross-model skill transferability, pass rate (%).

Figure 6: Per-domain pass rates on SkillsBench. Three conditions are compared using Claude Opus 4.6: no-skill baseline, human-curated skills, and CoEvoSkills self-evolved skills, across 11 professional domains. Numbers in parentheses indicate task counts. Self-evolved skills outperform human-curated skills in 9 of 11 domains. The arrow highlights Natural Science, where human-curated skills degrade performance, whereas self-evolved skills yield substantial gains, evidencing human–machine cognitive misalignment.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Pass rate (%)	$\Delta$ vs. Full
CoEvoSkills (Full framework)	71.1	—
W/O surrogate verifier	41.1	$-30.0$
W/O evolution (Context only)	42.4	$-28.7$
No-Skill Baseline	30.6	$-40.5$

Appendix

Table A1: Ablation results on SkillsBench, pass rate (%). Claude Opus 4.6 + Claude-Code. Ablation rows are single runs due to computational cost.

Figure B1: Distribution of verification cycles (left) and Ground Truth Oracle rounds (right) across 85 evolution tasks. Verification cycles include all host interventions ( Surrogate Verifier failures, Ground Truth Oracle evaluations). Ground Truth Oracle rounds count only the subset where the surrogate verifier passed and the oracle verifier was invoked. Failed tasks (red) cluster at higher iteration counts.

Round	Exit condition	Surrogate Verifier	Ground Truth Oracle	Key event
1	Verifier fail	0/15 (0%)	—	Initial skill has bugs
2	Checklist fail	15/15 (100%)	—	Progress checklist incomplete
3	Verifier pass	15/15 (100%)	3/4 (75%)	Period precision insufficient
4	Verifier pass	20/20 (100%)	3/4 (75%)	Precision fixed, alias check fails
5	Verifier fail	19/22 (86%)	—	Surrogate Verifier catches regression
6	Verifier pass	22/22 (100%)	4/4 (100%)	All tests pass

Appendix

Table C1: Evolution trace for the exoplanet transit detection skill. Each round records the trigger type, Surrogate Verifier outcome, Ground Truth Oracle outcome, and key event.

Ver.	Detrending	Algorithm	Precision strategy	Ground Truth Oracle
V1	Biweight	BLS	None	—
V2	Biweight (opt.)	BLS	None	75%
V3	Median filter	BLS	None	75%
V4	Savitzky-Golay	TLS	Two-stage + alias	100%

Appendix

Table C2: Skill versions for the exoplanet transit detection task. Each row records the detrending method, search algorithm, precision strategy, and Ground Truth Oracle outcome.

Aspect	Human-curated (5 skills)	Self-evolved (1 skill)
Total size	1,096 lines across 5 SKILL.md	64 lines SKILL.md + 142 lines Python
Executable code	None (documentation only)	9 callable functions
Algorithm guidance	Lists BLS, TLS, Lomb-Scargle equally	Prescribes TLS with justification
Period refinement	2-line tip: “refine candidates”	Two-stage search: broad then $\pm$ 2% narrow
Alias detection	“Check for aliasing” (1 sentence)	Function testing $P$ , $2P$ , $P/2$ automatically
Precision handling	Not addressed	Enforces 5-decimal output formatting

Appendix

Table C3: Structural comparison between human-curated and self-evolved skills for the exoplanet transit detection task.

LLM Agent Skill Learning

Large language model agent skill learning studies how agents acquire, represent, reuse, and refine procedures for multi-step tasks.

See all 379 papers

Skills-Coach: A Self-Evolving Skill Optimizer via Training-Free GRPO
Apr 30, 2026Yu Tian, Jiawei Chen, Lifan Zheng +5LLM Agent Self-Improvement LLM Agent Skill Learning
University of Chinese Academy of Sciences · 2East China Normal University · 3Zhongguancun Academy +3
Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills
Jul 24, 2026Siyuan Huang, Pengyu Cheng, Haotian Liu +10RL for Language Model Reasoning LLM Agent Skill Learning
Qwen Large Model Application Team, Alibaba · The Chinese University of Hong Kong · Renmin University of China +5
MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
May 26, 2026Huawei Lin, Peng Li, Jie Song +2LLM Agent Skill Learning LLM Agent Skill Retrieval
ByteDance Inc.

Apr 30, 2026J/K move · Enter open · S save

Skills-Coach: A Self-Evolving Skill Optimizer via Training-Free GRPO

Yu Tian, Jiawei Chen, Lifan Zheng, Mingxiang Tao+4

University of Chinese Academy of Sciences · 2East China Normal University · 3Zhongguancun Academy+3

LLM Agent Self-Improvement LLM Agent Skill Learning Prompt Optimization

Open paper AskPDF arXiv

We introduce Skills-Coach, a novel automated framework designed to significantly enhance the self-evolution of skills within Large Language Model (LLM)-based agents. Addressing the current fragmentation of the skill ecosystem, Skills-Coach explores the boundaries of skill capabilities, thereby facilitating the comprehensive competency coverage essential for intelligent applications. The framework comprises four core modules: a Diverse Task Generation Module that systematically creates a comprehensive test suite for various skills; a Lightweight Optimization Module dedicated to optimizing skill prompts and their corresponding code; a Comparative Execution Module facilitating the execution and evaluation of both original and optimized skills; and a Traceable Evaluation Module, which rigorously evaluates performance against specified criteria. Skills-Coach offers flexible execution options through its virtual and real modes. To validate its efficacy, we introduce Skill-X, a comprehensive benchmark dataset consisting of 48 diverse skills. Experimental results demonstrate that Skills-Coach achieves significant performance improvements in skill capability across a wide range of categories, highlighting its potential to advance the development of more robust and adaptable LLM-based agents.

Abstract

LLM Agent Skill Learning

Explore similar work

Skills-Coach: A Self-Evolving Skill Optimizer via Training-Free GRPO

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation