Skill∞: Recursive Self-Improvement for Agent Skill Optimization

Skill∞ enables recursive self-improvement in agent skill optimization: a meta-agent optimizes task skills following the same meta-skill process, and uses the same process to optimize the meta-skill itself, with agents leading throughout. Meta-skills and task skills are decoupled and can be reused across domains. On 83 SkillsBench tasks, Skill∞ achieved a 60.7% pass rate, higher than the 56.4% of manually curated skills and the 54.5% of SkillEvolver.