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

The framework known as CoEvoSkills, outlined in arXiv paper 2604.01687, empowers AI agents to independently create intricate skill packages that consist of multiple files. The term 'skills,' introduced by Anthropic for LLM agents, pertains to complex, multi-step tasks that basic tool functions cannot manage. Unlike standalone tools, skills consist of interconnected multi-file components. Current methods for generating skills require extensive labeling and often face cognitive misalignment between humans and machines, negatively impacting agent effectiveness, as indicated by SkillsBench assessments. CoEvoSkills integrates a Skill Generator that refines skills iteratively with a co-evolutionary verification system, enabling agents to autonomously develop skills without human assistance. This framework aims to address the challenges posed by existing self-evolving methods for tools, which are unsuitable for the intricate nature of skills.