A8gent
GlossaryAgent concepts

What is planner-executor?

Also called: plan and execute, planner executor agent

Planner-executor is an agent architecture that splits the work in two: one model call produces an explicit plan, then a cheaper executor carries out each step in order.

Written by A8gent
Contents
  1. At a glance
  2. What planner-executor means
  3. How the two halves are wired
  4. Cost and where the savings come from
  5. What you should know
  6. FAQ
  7. Related terms

At a glance

Category
Core agent concept
Also called
Plan and execute
Planner
One call, larger model, returns a structured ordered plan
Executor
Runs each step, often on a smaller model or no model at all
Key advantage
The plan is an artefact you can log, price and approve before anything runs
Key risk
Frequent replanning erases the cost advantage

What planner-executor means

Planner-executor splits an agent into two roles. A planner reads the goal once and produces an explicit, ordered plan - a list of steps with the tool and inputs for each. An executor then runs those steps, usually with a smaller and cheaper model, or with no model at all where the step is deterministic. It is the alternative to ReAct, which decides each step only after seeing the result of the last.

The commercial appeal is auditability. Because the plan exists as an artefact before anything runs, you can log it, show it to a human for approval, estimate its cost, and refuse it if it contains a step you did not authorise. With ReAct there is no plan to inspect - you find out what the agent intended by watching what it did.

How the two halves are wired

The planner gets the goal, the available tools and any constraints, and is asked for structured output: a JSON array of steps. Because this is a single call with a schema, it is a natural fit for function calling in structured mode. The executor iterates that array, calling tools and collecting results.

Real systems add a third piece: replanning. When a step fails or returns something the plan did not anticipate, the executor stops and hands the partial state back to the planner for a revised plan. Without replanning, planner-executor is brittle - the plan was written before anyone knew what the data looked like. With it, you get most of ReAct's adaptability at a fraction of the calls, because replanning happens on exception rather than every step.

Cost and where the savings come from

The saving is real but specific. A ReAct run makes one expensive model call per step, each re-reading a growing transcript. A planner-executor run makes one expensive planning call plus N cheap execution calls that only carry the step and its inputs. On a five-step task with a frontier planner and a small executor, that is routinely a large reduction in spend, and it also cuts latency because execution steps are short.

The savings vanish if you replan on every step, which is just ReAct with extra ceremony. Measure the replan rate: if more than a small share of runs replan more than once, the task is not plannable and you should use ReAct.

What you should know

Planner-executor suits tasks where the shape is known and the details vary - onboarding a client, reconciling a statement, preparing a quote from a job spec. It suits human approval especially well, because a plan is something a person can read in ten seconds where a transcript is not.

It is a poor fit for open-ended investigation, where the second step genuinely depends on what the first returned. It is also a poor fit when tools are unreliable, because every failure triggers a replan. In both cases the honest answer is a capped ReAct loop with good observability, or a plain agentic workflow if the branching is small enough to draw.

Planner-executor FAQ

What is a planner-executor agent?

An agent split into two roles: a planner that reads the goal once and writes an explicit ordered plan, and an executor that carries out each step. It contrasts with ReAct, where the agent decides each step only after seeing the previous result.

Why is planner-executor cheaper than ReAct?

ReAct makes one large model call per step, each re-reading the whole growing transcript. Planner-executor makes one large planning call and then a series of small execution calls that carry only the step and its inputs. On multi-step tasks with a small executor model, the difference is substantial.

What happens when a step in the plan fails?

You hand the partial state back to the planner for a revised plan. That replanning step is what stops the pattern being brittle, since the original plan was written before anyone had seen the real data. Cap the number of replans so a failing tool cannot cause an endless plan-fail-replan cycle.

When should I not use planner-executor?

When the second step genuinely depends on interpreting what the first returned, or when your tools fail often enough that most runs replan. In both cases you are paying planning overhead for no adaptability benefit, and a capped ReAct loop is the simpler answer.

Does planner-executor make human approval easier?

Considerably. A plan is a short, ordered list a person can read and approve or reject in seconds, before any action is taken. With ReAct there is nothing to approve up front, so the human checkpoint has to sit in front of each risky tool call instead.

Can I build planner-executor without code?

Partly. You can implement it in n8n or Make by having an LLM node return a JSON array of steps and then looping over that array with a Split In Batches or Iterator node. Replanning is the part that gets awkward visually, and it is usually where teams move to a code-based framework.

Primary source

LangGraph - agent control and human-in-the-loop concepts

Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.

Want this built rather than explained?

Knowing what planner-executor means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.