SkillOpt: Executive Strategy for Self-Evolving Agent Skills

Researchers introduced SkillOpt, a text-space optimizer that edits agent skill documents using scored rollouts and keeps only changes that improve validation.

Across six benchmarks, seven models, and three agent harnesses, SkillOpt matched or beat competing methods and showed strong transfer across settings.

The approach treats skills as external state, making updates reproducible and improving benchmark performance without adding deployment-time inference cost.