← Research library

The planner-executor pattern: separating what to do from doing it

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

As marketing agents take on multi-step campaigns, one design choice shapes everything after it: whether the same agent decides what to do and then does it. For consequential work, we recommend keeping those two jobs apart.

One of the most useful patterns in agent design is also one of the simplest: separate planning from execution. A planner agent takes a goal and produces a plan, a list of concrete tasks with dependencies. Executor agents, or people, carry out those tasks. The planner never touches customer-facing tools, and the executors never decide strategy.

This separation mirrors how good marketing teams already work. A campaign lead writes the plan and specialists execute it. Applying the same structure to agents brings the same benefits.

Why separate planning from doing

The first reason is that plans can be reviewed before anything happens. A plan is a cheap artifact. A person can read it in a few minutes, adjust it, and approve it before any email is drafted or any budget is spent. Reviewing a plan is far easier than reviewing the scattered results of an agent that planned and acted at the same time.

The second is that executors can be simpler and safer. An executor with one job, such as drafting an email from a given brief, needs fewer tools and less context. Narrow agents are easier to test and less likely to wander into work nobody asked for.

The third is that failures are easier to locate. When a campaign goes wrong, you can ask whether the plan was bad or the execution was bad. In a combined agent, those two causes are tangled together, and fixing one can quietly break the other.

What a good plan contains

A plan should be structured so that both people and executors can use it. We suggest these fields as a minimum.

  • Goal and success measure: what the plan is for and how you will judge it.
  • Tasks: each with an owner type (agent or person), inputs, expected output, and deadline.
  • Dependencies: which tasks must finish before others can start.
  • Assumptions: what the planner took as given without confirming.
  • Risks and review points: where the planner recommends a person check the work.

If the planner cannot fill in these fields, the goal is probably underspecified. That is worth knowing before work starts, and a planner that says so is more useful than one that fills the gaps with guesses.

Each task in the plan becomes a handoff when the hub routes it to an executor. The fields above map directly onto the handoff contract we describe in designing agent handoffs. The task's inputs are the payload, the goal and constraints are the context, and the expected output sets the acceptance criteria.

Where plan review sits

Plan review is step three in our orchestration run diagram, after the trigger and the plan and before any delegation. For consequential work, a person reviews and adjusts the plan there, because plans are cheap to change and executed work is not.

Not every plan needs the same scrutiny. A routine weekly report plan might be approved automatically once the planner has a track record. A plan that commits budget, changes pricing, or reaches a new audience should go to a person every time. Our autonomy ladder describes how to decide which plans earn lighter review.

A worked example

Consider a hypothetical product launch with a goal of booking demos from mid-market accounts over four weeks. The planner produces 12 tasks. They cover audience definition, a research brief, a landing page draft, three emails, two paid social variations, a sales enablement note, tracking setup, a consistency review, and a launch checklist.

The plan lists one assumption: that the existing mid-market segment is current. It flags two risks: the landing page needs legal review for a comparison claim, and the paid budget has not been confirmed. A marketing lead reads the plan in a few minutes, corrects the segment assumption, removes the comparison claim, and confirms the budget.

Only then does the hub route the tasks to executors. The email agent drafts from the corrected segment, the landing page agent never writes the comparison claim, and the paid agent works within the confirmed budget. Three problems were fixed in a plan, at almost no cost, instead of in finished assets.

Replanning

Plans meet reality and change. A good system lets executors report problems back, such as missing data, failed checks, or unexpected results, and lets the planner revise the remaining tasks.

The key rule is that replanning should be visible. If the plan changes materially, the revised plan goes back for the same review the original received. A planner that quietly rewrites its own plan midway through a campaign has undone the main benefit of the pattern.

We suggest defining "material" in advance. A change to budget, audience, offer, claims, or timing is material. A change to the order of two independent drafting tasks usually is not. Writing the line down keeps the planner from deciding it case by case.

Failure modes to watch

The pattern has its own failure modes. We suggest watching for these in the first few weeks.

  • Overplanning: the planner breaks simple work into many small tasks, adding handoffs and delay.
  • Vague tasks: tasks lack clear inputs or outputs, so executors guess.
  • Hidden replanning: the plan changes without a new review.
  • Executor strategy: an executor reinterprets its task and makes a decision the plan did not authorize.

Executor strategy is the subtlest. It often looks like initiative, such as an email agent choosing a different offer because it seemed more relevant. The fix is to make the executor return the suggestion to the planner rather than act on it.

When the pattern is overkill

For small, single-step tasks, a planner adds overhead without much benefit. The pattern earns its place when the work involves several distinct steps or specialists and those steps depend on each other. It also needs mistakes to be expensive enough that reviewing a plan first is worthwhile.

A weekly report probably does not need a planner. A multi-channel product launch almost certainly does.

How it fits in the hub

In our Agentic Hub architecture, the planner sits inside the hub and turns goals into tasks, while specialist agents handle execution. The hub's approval and state tracking then make sure executors only act on tasks from an approved plan.

We are studying how to measure plan quality directly. One candidate is the number of changes people make during plan review, tracked over time. A falling count suggests the planner is learning the team's constraints, and we want to know whether it predicts fewer problems in execution.

How to apply this: for your next multi-step campaign, ask an agent to produce only a plan in the structure above. Review it as a team before any execution starts, and count the changes you make. That count tells you how much autonomy the planner has earned.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.