Plan-and-execute
A planner model writes a step-by-step plan, executor models carry out the steps, and the planner revises the plan when a step fails.
Plan-and-execute splits an agent into two roles. A planner model writes a step-by-step plan for the whole task. One or more executors carry out the steps, calling tools. When a step fails or the results change the picture, the planner is called again to replan or to give the final answer.
The LangChain team’s post Plan-and-Execute Agents (February 2024) lists the benefits: the big model is not consulted after every action, smaller and cheaper models can run the steps, and making the planner think through all the steps may improve the result. The idea builds on research such as Plan-and-Solve prompting (Wang et al., 2023). The cost is rigidity. A plan written before the first tool call rests on guesses, and a plan that is never revised fails at the first surprise. Replanning is the part to get right.
The usual alternative is the ReAct-style agentic loop, where the model picks one step at a time after seeing each result. That adapts better, but it costs a model call per step and can wander. Plan-and-execute sits between a fixed agentic workflow and a free agent. The plan is also a natural review point: show it to a person before execution (human-in-the-loop). For a strong model that only advises a cheaper one at key moments, see advisor model.
Sources
- LangChain: Plan-and-Execute Agents langchain.com
- arXiv: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models arxiv.org
Related terms
- Agentic loop: The cycle of model call, tool call and result that an agent repeats until it stops.
- Agentic workflow (workflow vs agent): A process where your code fixes the steps and a model does some of them, as opposed to an agent that decides its own steps.
- Orchestrator-workers pattern (supervisor): One model plans the task, hands parts of it to worker models or subagents, and combines their results.
- Advisor model: A cheaper executor model does the work and consults a stronger advisor model at key moments, such as planning or a final review.
- Model routing and fallback: Sending each request to the model that fits it, and to another one when the first fails.
- Human in the loop: A person approves, edits or rejects the agent’s action before it takes effect.
- Cost per task (agent unit economics): What one finished agent task really costs: tokens, retries, tool calls and the human time to review it, compared with doing it another way.
