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.
Continue to AI University →