EvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (April 2026) logo

EvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (April 2026)

Free

Skill Generator iteratively refines agent skills while a Surrogate Verifier co-evolves to provide actionable feedback without ground-truth; surpasses human-written skills on SkillsBench in 5 rounds; works on Claude Code and Codex

FreeFree tier
Type
Open Source

About EvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (April 2026)

CoEvoSkills (Self-Evolving Agent Skills via Co-Evolutionary Verification) is a research framework that enables LLM agents to autonomously generate complex, multi-file skill packages for multi-step professional tasks. It couples a Skill Generator that iteratively refines skills with a Surrogate Verifier that co-evolves to provide informative and actionable feedback without requiring ground-truth test content. On the SkillsBench benchmark, CoEvoSkills achieves the highest pass rate among five baselines on both Claude Code and Codex, and demonstrates strong generalization capabilities to six additional LLMs. The approach addresses the label-intensive nature of manual skill authoring and mitigates human–machine cognitive misalignment, offering a path toward fully autonomous skill development for agentic AI systems.

Key Features

Enables agents to autonomously construct complex, multi-file skill packages
Couples Skill Generator that iteratively refines skills with Surrogate Verifier that co-evolves
Provides informative and actionable feedback without access to ground-truth test content
Achieves highest pass rate among five baselines on SkillsBench
Demonstrates strong generalization to six additional LLMs beyond Claude Code and Codex
Addresses label-intensive manual authoring and human–machine cognitive misalignment

Pros & Cons

Pros
  • Highest pass rate on SkillsBench compared to five baselines
  • Self-evolving without requiring ground-truth or human-labeled data
  • Generalizes well across multiple LLMs (Claude Code, Codex, 6 others)
  • Addresses the complexity of generating multi-file skill packages
  • Open-source framework with code to be released

Best For

Enabling LLM agents to autonomously generate skills for multi-step professional tasksImproving agent performance on complex benchmarks like SkillsBenchReducing reliance on manual skill authoring in agentic AI systemsAutomated refinement of agent capabilities through co-evolutionary feedback loops

FAQ

What is CoEvoSkills?
CoEvoSkills is a self-evolving skills framework for LLM agents that autonomously constructs complex, multi-file skill packages. It uses a co-evolutionary approach with a Skill Generator and Surrogate Verifier, achieving state-of-the-art results on the SkillsBench benchmark.
How does CoEvoSkills differ from existing skill generation methods?
Unlike manual authoring or simple tool-generation approaches, CoEvoSkills generates structured multi-file skill bundles and uses a co-evolving surrogate verifier to provide feedback without ground-truth test content, enabling iterative refinement.