All tools
Evals & observabilityUpdated 2 days ago

SkillOpt

By Microsoft Research

MIT-licensed Python framework from Microsoft Research, around 16,800 GitHub stars, that treats an agent skill file as a trainable parameter: a frozen model runs scored task trajectories, an optimizer model proposes add, delete, and replace edits in minibatches, and a change survives only if it improves a held-out validation split, leaving one deployable best_skill.md of roughly 300 to 2,000 tokens.

Best for

  • tuning agent skill files against measured results instead of by hand
  • raising agent accuracy without touching model weights
  • validation-gated edits that keep a record of rejected changes
  • overnight offline refinement drawn from past agent sessions
  • skills that deploy into Claude Code, Codex, and Copilot
Recent changes
  • Sep 9, 2026Added to the evals category: Microsoft Research framework that optimizes an agent skill file through scored trajectories, minibatched add/delete/replace edits, and a held-out validation gate, shipping one best_skill.md artifact with no weight update.

Other Evals & observability