Microsoft research shows agent 'skills' transfer across AI tools

Microsoft's SkillOpt research found that an optimized skill file trained on one AI coding tool can improve a completely different one. A SpreadsheetBench skill trained using Codex raised Claude Code's score from 22.1 to 81.8, slightly above the 80.4 Claude Code reached training its own skill. Transfer varied sharply by task, retaining 102% of the benefit on spreadsheet tasks but only 10% on math tasks.

Why it matters: If reusable skill artifacts can move between competing coding agents, it undercuts the idea that each vendor's agent needs bespoke fine-tuning to get good at a task, pointing toward a shared ecosystem of portable agent capabilities instead of walled gardens. The sharp variance between spreadsheet and math tasks also hints that today's skill-transfer techniques exploit surface pattern-matching more than genuine reasoning transfer.