SkillOpt: Microsoft Shows Agent Skills Can Be Auto-Optimized via Text Gradients
Microsoft Research and Shanghai Jiao Tong University released SkillOpt, a framework that treats agent SKILL.md files as learnable parameters. An optimizer LLM reads training trajectories from a frozen target model, proposes bounded edits to the skill file as "textual gradients," and validates changes via sandbox execution. Applied to GPT-5.5 on search question-answering, the approach lifted scores from 77.7% (no skill) to 87.3%, beating human-written and GaiPa baselines. Final winning skills are tiny: 1–3 atomic edits, approximately 880 tokens. Open-sourced as SkillLens on GitHub.
Why It Matters
Automated optimization of procedural knowledge documents creates a path to self-improving agent instructions without model retraining. SkillOpt-optimized skills are portable across Claude Code, Codex, and Cursor — establishing a cross-platform transferable procedural intelligence layer.