SkillOpt: Agent Skills as Trainable Parameters

Microsoft Research presents SkillOpt, a method that treats an agent’s skill file as a trainable parameter outside a frozen target model. It replaces one-shot manual instruction editing with a controlled optimization loop using bounded text edits, validation gating, rejected-edit feedback, and slower meta-updates. The reported evaluation covers six benchmarks, seven target models, and three execution modes. SkillOpt was best or tied-best in all 52 evaluation cells while leaving model weights unchanged. The authors also report transfer across model scales, agent harnesses, and related tasks, indicating that the learned skills may encode reusable workflow knowledge rather than benchmark-specific prompt changes.