Microsoft's SkillOpt Demonstrates Optimized Agent Skills Transfer Across Model Sizes and Between Code Harnesses

Researchers from Microsoft and three Chinese universities have developed SkillOpt, a text-space optimizer that trains natural-language skill documents while keeping target models frozen. An optimizer model proposes bounded add-delete-replace edits based on scored rollouts, accepting only edits that strictly improve performance. Transfer experiments show skills trained on larger models retain meaningful gains when deployed on smaller variants—SpreadsheetBench on GPT-5.4-mini preserves 82% of gains originally achieved on GPT-5.4. Notably, a skill optimized within Codex lifted Claude Code's performance from 22.1 to 81.8, slightly exceeding Claude Code's own in-domain optimization result of 80.4, suggesting procedural skills transfer more effectively than reasoning-heavy ones across different execution environments.