Recipe Overview

Large tasks often need an explicit roadmap. A plan-and-execute agent first generates a structured plan, then carries it out. The problem it solves is redundant LLM reasoning. In this approach, one part of the system (often a larger model) produces an outline of steps, then smaller models or tools execute them. For instance, for 'analyze this dataset,' the agent might plan: 1) load data, 2) check for missing values, 3) generate summary statistics, 4) create visualizations. This separation allows optimization: planning can use expensive models while execution uses faster, cheaper ones.