What is a Plan-and-Execute (Agent Pattern)?

It contrasts with a purely reactive agent loop (like plain ReAct-style step-by-step reasoning) where the model decides only its very next action with no explicit forward plan. In plan-and-execute, a planning step produces an ordered list of sub-tasks up front — "search for the vendor's API docs," "extract the auth flow," "write the integration code," "run the tests" — and an executor then works through that list, calling tools as needed for each sub-task. The advantage is that the model reasons about the whole task's structure before diving in, which tends to produce more coherent multi-step behavior and avoids getting stuck repeating small reactive steps without ever stepping back. The tradeoff is that a rigid plan can become stale: if step three reveals the original plan's assumption was wrong, a good implementation needs a re-planning step, not blind adherence to the original list. Many production agent frameworks use a hybrid: plan-and-execute for the outer loop's structure, with a more reactive, tool-calling loop inside each individual planned step, giving both the coherence of upfront planning and the adaptability of moment-to-moment reasoning.