SkillOpt: self-evolving agent skills
Microsoft Research's SkillOpt improves an agent's instructions automatically through trial and error, like gradient descent for text, with no extra cost at deployment.
Writing good instructions for an AI agent is mostly manual work. We draft a skill or a system prompt, watch where the agent fails, and rewrite by hand. SkillOpt, from Yifan Yang and colleagues at Microsoft Research, automates that loop. It treats an agent’s natural-language instructions the way gradient descent treats weights: it runs the agent, finds what is not working, and makes small targeted edits until performance improves.
How it works
- Test. Run the agent on a task and observe where it falls short.
- Diagnose. Identify which parts of the instructions are responsible for the failures.
- Edit. Make small, targeted changes to the skill text, then measure whether the change helped.
The loop repeats, so the skill evolves through trial and error rather than from a single hand-written draft. Because all of the optimization happens before deployment, there is no extra inference cost when the improved skill is used in production.
Results
Across six benchmarks and seven models, SkillOpt improved accuracy by up to 24.8 points, with no added cost at deployment.