Microsoft’s open-source SkillOpt automatically upgrades AI agent skills without touching model weights

Microsoft released SkillOpt, an MIT-licensed framework that treats AI agent skill documents as optimizable objects rather than static instructions. The system uses performance feedback to revise markdown-based skills, improving benchmark accuracy for models such as GPT-5.5 and Qwen without changing model weights.

SkillOpt points to a growing trend of improving enterprise AI through external, editable layers rather than expensive model retraining. If reliable, skill optimization could make agent deployments easier to maintain, more transferable across domains, and less dependent on manual prompt engineering.