🖥 Microsoft open-sources SkillOpt for agent skill tuning

Microsoft released SkillOpt, a framework that improves agent abilities by editing markdown skill files in the background. It runs a learning loop in text space, logging agent actions and suggesting small skill file updates after verification on a test set.

SkillOpt works by iterating on files, applying minor changes that pass safety rules and verification before adoption. This prevents large accidental regressions while boosting performance.

The framework shows consistent gains across models and benchmarks. On GPT-5.5, Codex and Claude Code saw average improvements of +21.8 and +18.6 respectively. The project includes code, overview, and usage instructions.

📊@tech