Tag
This paper argues that full-horizon planning with lazy replanning is more efficient than step-by-step execution for data-centric LLM agent tasks, using fewer tokens while maintaining accuracy.