arXiv cs.AIMonday · May 25, 2026FREE

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

SkillOpt, detailed in a new arXiv preprint (2605.23904), proposes a novel approach to agent skill optimization. Unlike current methods that rely on hand-crafted or one-shot generated skills, SkillOpt treats a skill document as external state that can be iteratively improved via a separate optimizer model. The optimizer applies bounded add/delete/replace edits based on scored rollouts, accepting edits only when they strictly improve a held-out validation score. Key components include a textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow/meta updates to ensure stability, all without adding inference-time model calls. The system was evaluated across six benchmarks (including diverse reasoning and coding tasks), seven target models (e.g., GPT-4, Claude, Codex), and three execution harnesses (direct chat, Codex, Claude Code). SkillOpt achieved best or tied performance on all 52 evaluated (model, benchmark, harness) cells, outperforming human-written skills, one-shot LLM-generated skills, and prior self-revision methods like Trace2Sk. The paper is available on arXiv and was published on May 25, 2026.