LLM agents handle many quasi-static tasks with semantics and geometry. Dexterous actions lack clear semantics and must be learned from physical feedback.
