Plan-and-execute
A canonical two-stage agent pattern: a planner LLM produces a structured multi-step plan, then an executor (often a cheaper model) carries out each step using tools.
Plan-and-execute splits agent work cleanly: the planner thinks hard about strategy once; the executor handles each tactical step. This contrasts with ReAct, which interleaves reasoning and action on every step. Plan-and-execute is faster, cheaper, and more debuggable for known task shapes.
The tradeoff is rigidity. If the world changes mid-task, a pure plan-and-execute agent will charge ahead with the stale plan. Modern implementations add a re-plan trigger: when an executor result diverges from expectation, the planner gets called again.
LangGraph, OpenAI Agents SDK, and CrewAI all ship plan-and-execute templates. It is the right starting pattern for any task where the steps are predictable but tedious — research reports, multi-file refactors, account audits.
Frequently asked
Plan-and-execute vs. ReAct — which should I choose?+
ReAct when the environment is dynamic and you need step-by-step reasoning (web browsing, code debugging). Plan-and-execute when the task is repeatable and you want cost predictability (research, scheduled workflows).
Can the planner and executor be the same model?+
They can, but production setups usually run a frontier model as planner and a cheap fast model as executor. The cost gap is 5–20×; the quality gap on individual tool calls is small.
