标签
This paper argues that representing agent skills as code, rather than natural language, yields the best cost reduction for LLM agents, and introduces SpeedRunner, a coding agent that learns programmatic skills from past trajectories to improve performance while cutting costs across embodied environments.