An agent architecture where a planner produces a full task breakdown up front, then cheaper or deterministic executors run steps—contrasted with ReAct, which interleaves reasoning and action each cycle.
Plan-and-Execute trades flexibility for efficiency: you plan once, parallelize where possible, and avoid re-invoking a large reasoning model on every micro-step. If the initial plan is wrong, replanning can be expensive—that is the main downside versus ReAct.
When teams choose it
Good fit for well-scoped workflows (deploy checklists, data pipelines) where step order is mostly known. Poor fit when the task is exploratory and the next action depends heavily on fresh observations.
See also
Hierarchical multi-agent systems for delegation across specialists.