LLM agents do much better when you parcellate work for them in advance. This kind of planning also opens up more economical downstream agential workflows such as using a local model to implement work that a more capable model from one of the frontier labs has planned.