Plan-and-Execute writes the full plan once and executes it in sequence; ReAct interleaves a Thought and Action at every step. The difference is when planning happens.
Imagine planning a road trip. One way is to map out every stop before you leave, then just drive between them without checking the map again. That is Plan-and-Execute. Another way is to drive a bit, look around, decide where to go next, drive a bit more. That is ReAct. The first is faster when the route is predictable because you only think hard once. The second adapts better when something unexpected shows up, like a closed road, because you re-decide with the new information. Real agents often blend the two: plan first, execute several steps, and only re-plan when the world surprises you. Both patterns turn a single model call into an agent loop, just with different rhythms.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Plan-and-Execute and ReAct are the two canonical agent loops, and a junior engineer should be able to state the difference in one sentence. Plan-and-Execute writes the full plan once at the start, then executes the steps in sequence. ReAct interleaves a Thought and an Action at every step. The structural difference is when the planning happens.
This explanation walks through what each pattern looks like in practice, why the timing of planning is the load-bearing distinction, where each pattern wins, and how production systems blend them with replanning triggers.
Plan-and-Execute: plan once, run many
A Plan-and-Execute agent has two distinct components and two distinct phases.
The planner is typically a frontier LLM call (Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro). It reads the user's goal, often along with available tool descriptions, and produces a structured plan. The plan is usually a numbered list: step 1 fetches data, step 2 transforms it, step 3 summarises the result. Each step names the tool and its inputs.
The executor then runs each step in sequence. The executor can be a separate cheaper model that just formats tool calls, a deterministic runner that executes pre-specified tool calls, or even a different specialised agent per step. The defining property is that the planner is not re-engaged between steps. Once the plan is written, the executor follows it.
This is a fundamentally different shape from a single chain-of-thought call. Plan-and-Execute still involves multiple model calls and a real loop; it just puts the heavy planning model in a different place. The planner call is expensive but happens once. The executor calls are cheap and happen many times.
The pattern works best when the steps are largely knowable in advance. Multi-step research summaries, scripted data pipelines, and structured workflows where the branch points are rare all fit cleanly into Plan-and-Execute.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- LangGraph's prebuilt plan-and-execute agent runs a planner node once to produce a step list, then loops an executor node with a conditional replanning edge that fires on plan-contradicting observations.
- CrewAI's hierarchical-process mode uses a manager agent that plans the task into subtasks and delegates each to a worker agent, with the manager only re-engaging when a worker reports failure.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy is Plan-and-Execute often cheaper than ReAct on a 12-step research task?
Count the planner-model calls. Plan-and-Execute: one frontier planner call plus 12 cheap executor calls. ReAct: 12 frontier planner calls. If the planner is a Claude Opus 4.7 or GPT-5.5 call, the difference is roughly 10-12x on planner-side spend.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating Plan and Execute and ReAct as the same loop with different names. They differ in when the planning step runs: once up front versus at every turn.
60 second bullets to scan on the way to the call.
Define Plan and Execute as planner once then executor.
Define ReAct as Thought and Action interleaved every step.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.