SkillNEO Uses Non-Elitist Evolution to Improve Agent Skills

Summary

Agent skills convert procedural knowledge into reusable artifacts that language-model agents can refine through execution feedback. The authors argue that conventional elitist evolution can discard versions with lower immediate rewards even when those versions contain partial repairs needed for later improvements. SkillNEO addresses this with a non-elitist process that separately chooses the version used for execution and the versions retained for further development. It compares related versions on the same task to identify steps worth preserving and failures that still need repair, and can revisit earlier versions when later evidence becomes available. Across 86 SkillsBench tasks in eight domains, including software engineering and office work, the method increased GPT-5.4 Mini's average reward from 49.1% to 66.4%. That result falls within the original-skill performance range reported for GPT-5.5 and GPT-5.6 Luna, and improvements appeared in all eight domains. The resulting skills also improved stronger models without additional editing. On the 41 tasks with modified skills, reuse raised Luna's performance by 10.5 percentage points and put GPT-5.5 10.6 points above its published original-skill reference. Execution records showed that partially unsuccessful Mini-generated revisions could still help stronger models, while search histories showed that some lower-scoring versions were eventually selected for use. The authors conclude that skill evolution can convert a weaker model's partial progress into reusable procedural knowledge for more capable agents.