Planner-Executor
Planner-Executor agent pattern: separating planning and execution phases for more reliable and debuggable AI agent workflows.
Last updated
After this section you can
- Choose between a linear, DAG and conditional plan, and estimate a DAG’s latency from its critical path
- Place a human gate on the plan and bind the executor to the approved version
- Decide between planner-executor, ReAct and the hybrid, and re-plan on failure with a cap
12
Planner-Executor: Divide & Conquer
ReAct decides one step at a time. A planner-executor agent writes the whole plan first, then runs it. That turns the plan into something you can validate, approve, parallelise and repair before a single side effect fires.
Key idea
Split the agent in two. A strong model writes a structured plan once; cheap models or plain code execute it. The plan is data: you check it, gate it, run independent steps in parallel, and on failure re-plan only what is left.
Think once, up front. Everything after the gate is cheap, parallel and checkable.