Microsoft Automatically Optimizes AI Skills Without Model Changes

Microsoft has released "SkillOpt," an open-source framework that automatically optimizes "skills" for AI agents. Provided under the MIT License, this tool's greatest feature is that it can continuously improve skill documents that define agent behavior without making any changes to the model's weights (parameters).

In AI agents, "skills" refer to groups of instructions written in markdown format (.md) text files. They include domain-specific heuristics, tool usage policies, output constraints, known error patterns, and other elements, functioning as an external interface for agents to handle complex business workflows. While there is an advantage in being able to customize behavior without changing model weights, optimization of skill documents themselves has relied on manual work until now.

Traditionally, skill improvement has been a "guesswork" process where staff manually rewrites files and tries different changes through trial and error to see which improvements in performance. Yifan Yang, Senior Research SDE at Microsoft Research Asia, explains this challenge as follows: "The problem is not whether skills can be changed, but that changes cannot be guaranteed to be improvements. Without step size control, skills gradually deteriorate; without validation, seemingly harmless modifications secretly degrade performance; and without records of failed edits, the same mistakes are repeated——these three failure patterns occur repeatedly."

SkillOpt addresses these challenges by applying deep learning optimization techniques to text documents. It treats skill .md files as trainable objects and systematically updates their content based on performance feedback obtained from agent execution results. By introducing mathematical rigor to text optimization, it prevents disorderly modifications and achieves a mechanism that continuously validates whether changes actually lead to performance improvements.

In evaluations across various industry benchmarks, SkillOpt demonstrated significantly superior results compared to existing methods. Notable accuracy improvements were confirmed with models such as GPT-4.5 and Qwen, and the optimized skills, as "compact and transferable artifacts," can be easily applied to new domains. As AI agent practical deployment accelerates, SkillOpt, which can enhance agent expertise without modifying the model itself, appears to be a promising option for enterprises incorporating AI into their operations.