Planner-Executor
Planner-Executor agent pattern: separating planning and execution phases for more reliable and debuggable AI agent workflows.
Last updated
Planner-Executor: Divide & Conquer
Decide the whole plan first, then execute it. Separating the “what” from the “how” buys you an inspectable, editable, parallelizable plan — and a natural place to put a human before anything irreversible happens.
ReAct decides one step at a time, re-reasoning over the full history on every call. That is flexible but expensive, drift-prone, and opaque — you can’t see where it’s going until it gets there. Planner-Executor splits the agent in two: a planner (strong model, called once) produces a complete, structured plan; an executor (cheap model or plain code) runs each step. The plan becomes a first-class artifact — you can read it, diff it, edit it, gate it, and parallelize it before a single side effect fires. You trade ReAct’s step-by-step adaptivity for a plan you can reason about as a whole.