SkillOpt: Training Loop for Agent Skills

SkillOpt treats a natural-language skill document as the sole trainable state of a frozen agent and optimizes that document through an iterative propose-and-test loop. The frozen target model executes tasks to produce rollout evidence; a stronger optimizer model reflects on success and failure minibatches to propose structured add/delete/replace edits under an edit budget (a textual learning rate). Candidate skills are accepted only if held-out validation improves, while rejected edits are stored as negative feedback and a slow-update meta-skill provides longer-horizon memory. Bounded edits, gated selection, rejected-edit buffering, and slow updates prevent destructive rewrites and stabilize learning; the pipeline records messages, tool calls, verifier feedback, and scores to drive focused, reusable procedural rules rather than weight updates.

Empirically, optimizing skills in text-space yields large, consistent gains across multiple target models (GPT and Qwen families) and execution harnesses (direct chat, Codex, Claude Code), outperforming strong baselines on six benchmarks and matching or leading in every tested setting. Runs show dramatic per-benchmark improvements (e.g., spreadsheet and code harness boosts) and successful transfer: exported single-file skills transfer across model sizes (+15.2), harnesses (+31.8), and enable self-optimization (+10.4). Ablations confirm each optimizer component contributes materially, and deployment requires only the compact exported skill, not optimizer state.