Microsoft Researchers Reveal Breakthrough AI Tool for Portable Agent Skills

A team of researchers from Microsoft and several universities has developed a text-space optimizer called SkillOpt. This tool enables the transfer of optimized agent skill artifacts across different model scales and between various code harnesses, such as Codex and Claude Code.

SkillOpt works by training a single natural-language skill document while keeping the target model frozen. An optimizer model then reads scored rollouts and proposes bounded add/delete/replace edits. The exported artifact is one file, best_skill.md, which can be used across different environments.

The researchers tested SkillOpt on several benchmarks, including SpreadsheetBench and LiveMath. They found that procedural spreadsheet skills move well between models and code harnesses, but math-reasoning skills are less portable. One notable result was a Codex-trained SpreadsheetBench skill scoring 81.8 inside Claude Code, surpassing the latter's own in-domain result of 80.4.

The study demonstrates that portability is possible with SkillOpt, and the results have significant implications for AI development. Training cost can be paid once offline, and the optimized artifact can be deployed across various environments without additional inference-time calls.