Agent skills turn procedural knowledge into reusable artifacts, enabling large language model (LLM) agents to accumulate and refine expertise through experience. Building on this, skill evolution iteratively improves skills using execution feedback. However, genuine improvement often requires several interdependent edits, so low‑reward intermediate versions can still be valuable. Insisting on immediate reward gains would discard useful partial repairs before their benefits materialize.
Non‑elitist evolution permits new versions that score below existing ones to continue receiving edits, preserving potential improvements. For agent skills this demands evidence of which steps to keep and which failures to fix. We introduce SkillNEO (Non‑Elitist Evolutionary Optimization for Agent Skills), which separately selects a version for actual use and versions to keep improving. By comparing how related versions perform the same task, SkillNEO identifies working steps to preserve, remaining failures to address, and can revisit earlier versions using later evidence.
Across 86 SkillsBench tasks spanning eight domains (including software engineering and office work), SkillNEO raises GPT‑5.4 Mini’s average reward from 49.1% to 66.4%, entering the performance range of GPT‑5.5 and GPT‑5.6 Luna’s original‑skill baselines. Mini executes tasks, proposes revisions, and edits skills; final evaluation uses the same task instances. Gains appear in all eight domains. The same skills improve stronger models without further editing. On the 41 tasks with modified skills, reuse boosts Luna by 10.5 percentage points and places GPT‑5.5 10.6 points above its published original‑skill reference. Execution logs show these skills can benefit stronger models even when Mini still fails some task requirements. Search histories also reveal lower‑scoring versions that later lead to the finally selected skills.
Together, these findings demonstrate that skill evolution can turn a weaker model’s partial progress into reusable procedural knowledge, thereby enhancing more capable agents.
Review