Microsoft Research introduced SkillOpt, a method that trains agent skills by treating a SKILL.md file as a trainable external parameter instead of fine-tuning model weights. The approach keeps the target model frozen and iteratively edits the skill text to improve held-out performance.
How it works
- The frozen target model executes tasks and records scored trajectories.
- A separate optimizer model proposes structured textual edits to SKILL.md.
- Proposed edits are accepted only when they improve performance on a held-out validation set.
The system mirrors a traditional training loop: rollouts act as forward passes, reflection steps correspond to backward passes, and a textual edit budget functions like a learning rate to prevent destructive rewrites.
Evaluation and results
SkillOpt was evaluated across 6 benchmarks and 7 models, covering 52 experimental settings. It achieved the best or tied-for-best performance in every setting tested.
Key quantitative outcomes:
- With Claude Code as the harness running a GPT-5.5 target model, SkillOpt-trained skills boosted performance by an average of 18.6 points across benchmarks; Spreadsheet tasks saw a +58.3% increase.
- Cross-model and cross-harness transfer: a skill trained with Codex transferred directly to Claude Code and produced a +31.8% gain on SpreadsheetBench.
- Cross-version transfer: a skill trained on GPT-5.4 transferred to GPT-5.4-nano and still produced a +15.2% improvement.
- Self-optimizer mode: even when the target model serves as its own optimizer, the constrained, validated update loop discovers useful edits.
Operational details and safety
The full optimization process outputs a single best_skill.md file. At deployment, the target model never has access to the optimizer’s memory, rejected edits, or training state — only the final skill file is used.
License and availability
The entire project is open-sourced under the MIT license.
Why this matters
By treating agent behavior as editable text, SkillOpt enables reusable, auditable, and portable skills that can be transferred across models and execution harnesses without changing model weights. That promises practical benefits for agent development workflows, including easier versioning and inspection of behaviour.
