Microsoft and researchers develop SkillOpt to train AI agent instruction documents like model weights

Summary

Microsoft and three Chinese universities developed SkillOpt, a method that optimizes instruction documents (Markdown files) for AI agents using gradient-based training techniques traditionally applied to model weights. The approach boosts GPT-5.5's performance on procedural tasks by approximately 23 points, and the optimized skill files transfer effectively across different models and agent environments including Codex and Claude Code. This represents a novel approach to improving AI agent performance by treating instruction documents as trainable assets rather than static prompts.