Microsoft has released SkillOpt, a framework that allows developers to optimize agent skills with the same rigor as deep learning, but without touching model weights. The project treats the skill document itself as the trainable state of a frozen agent, using a separate optimizer model to apply bounded edits. This approach eliminates inference-time overhead while significantly boosting performance across multiple benchmarks.

Training Skills Without Weights

SkillOpt operates by turning scored rollouts into add, delete, or replace edits on a single skill document. A candidate edit is only accepted if it strictly improves a held-out validation score, mimicking the discipline of weight-space optimization. The system uses a textual learning-rate budget and a rejected-edit buffer to stabilize the training process, ensuring that the skill evolves reliably rather than degrading through loose self-revision.

Performance Metrics and Benchmarks

The results are substantial. On GPT-5.5, SkillOpt lifted average no-skill accuracy by 23.5 points in direct chat scenarios. Inside the Codex agentic loop, the improvement reached 24.8 points, while Claude Code saw a 19.1-point boost. The framework was evaluated across six benchmarks, seven target models, and three execution harnesses, achieving best or tied-best results on all 52 evaluated cells.

SkillOpt-Sleep and Integration

Version 0.2.0 introduces SkillOpt-Sleep, a nightly offline self-evolution engine available via the skillopt-sleep CLI. This tool harvests past sessions, mines recurring tasks, and consolidates validated skills behind a validation gate. The release also includes integration shells for major coding agents like Claude Code, Codex, Copilot, and Devin, along with an OpenClaw reference adaptation.

Key Takeaways

  • SkillOpt optimizes prompt skills, not model weights, reducing deployment costs.
  • GPT-5.5 accuracy improved by up to 24.8 points in agentic loops.
  • The framework is accepted as a Spotlight at NeurIPS 2026.

The Bottom Line

This is a game-changer for agent reliability. By treating prompt engineering as an optimization problem, Microsoft has solved the 'brittle skill' issue that plagues most autonomous agents.