Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Planner-executor architecture

  • Agents
  • Standard term
2 steps

A planner-executor architecture is an agent design in which one part, the planner, writes the list of steps, and a separate part, the executor, carries them out with tools and reports what happened.

The planner writes steps for the executor to carry out. Some designs follow that initial plan; others send results back to the planner so it can revise the remaining steps. Depending on the design, the executor gets only its current step or the whole plan for background. The roles can be separate models, separate prompts to one model, or a model paired with ordinary code.

Where you’ll meet it

A research agent is comparing health plans for your family. Its plan lists premiums, your doctors, and prescriptions as the things to check, and then it searches each insurer's site. The final table leaves out your child's pediatrician.

What to do

Compare the request, plan, search results, and table. If the pediatrician was never named, clarify the plan. If the name was included, find whether the lookup failed or its result was omitted from the table, and repair that step.

Easy to mix up

ReAct mixes reasoning and actions in one running thread, deciding each step as it goes. A planner-executor design writes the plan out first, so a separate part can follow it and revise it.

Check yourself

An agent writes its full plan before it acts. Which of these two jobs suits that design better?

Go deeperAn explicit plan creates a review point before execution

The split echoes classical AI planning, where a planner produces a sequence of actions and a separate system carries them out. With language models, a written plan also gives a point for review, since code or a person can check it before any step that spends money or can't be undone.

ReWOO, from Binfeng Xu and colleagues in 2023, plans tool work ahead of observations, using placeholders for results. It reported lower token use than the compared approach that repeatedly supplied tool history to the model. Tokens are the text units models process. The result concerns the tested design and tasks; other planner-executor systems may replan as observations arrive.

Practice Modern AI

Or take the learning quiz on a mix of every set