Study Maps When AI Agent 'Skills' Actually Help - and When They Break Down

A large controlled study of LLM agent 'skills' finds they work by anchoring messy reasoning into stable procedures, but breaks down the exact conditions that make them fail.