SkillOpt: Turning AI Agent Skills into Trainable Parameters
Microsoft Research introduces SkillOpt, a method that treats agent skill files as trainable parameters outside a frozen model, optimizing them through controlled text edits. Across 52 evaluation cells with 6 benchmarks, 7 models, and 3 execution modes, SkillOpt consistently outperforms baselines, achieving up to +23.5 points improvement without modifying model weights. Optimized skills transfer across model scales and agent harnesses, remaining compact and auditable.
Microsoft
OpenAI
Alibaba/Qwen
