Planner / Executor overview

By the end of this page you will know what this pattern does, when to reach for it, and what a member agent needs to be useful in one.

What it does

You give the system a team of AI Agents. For each turn, a planner reads the goal and writes a plan: which agents to run, in what order, and which of them can run at the same time. An executor runs that plan and composes the results into one answer.

  "Pull our Q3 numbers, research the competition, and write a
   launch email — then give me one summary"
                          │
                          ▼
                    ┌───────────┐
                    │  Planner  │  reads every member's skills
                    └───────────┘
                          │  writes a plan
          ┌───────────────┴───────────────┐
          ▼                               ▼
    Sales Analyst                  Market Researcher     ← run together
          └───────────────┬───────────────┘
                          ▼
                    Content Writer                       ← waits for both
                          ▼
                  Editorial Reviewer
                          ▼
                    one final answer

There is no order for you to author. Add or remove a member and the planner adapts, because it plans against whatever team it has at the time.

When to reach for it

Choose Planner / Executor when the work genuinely needs several agents and you cannot know in advance which ones, or in what order. A research request might need two sources this time and four the next; a planner handles that, and a drawn flow cannot.

The same team answers a narrower request with a smaller plan. “Research what competitors announced this quarter, then write a short blog post about how we compare” runs Market Researcher, then Content Writer, then Fact Checker, then Editorial Reviewer — one chain, no parallel step, because that is what the request needed.

Choose something else when:

• One specialist can answer each message on its own. Use the Conversation Router — it is cheaper and more predictable.
• You already know the sequence and want it every time. Use an Agent Graph. A planner that always produces the same plan is a graph you are paying an LLM to redraw.

What a member needs

The planner chooses agents by reading their skills — names, descriptions and example queries. That has one practical consequence worth acting on:

An agent with vague skills will not be chosen, or will be chosen for the wrong step.

So before adding an agent to a team, open it and check that each skill says plainly what it does. See Skills.

Members must be AI Agents with status Active. Task Agents cannot be members; expose one as a tool on a member agent instead.

What it does not do

• You cannot pin the order. If a step must always come after another, that is an Agent Graph.
• You do not connect members to each other. The catalog is a flat list; the plan supplies the structure.
• A plan belongs to one turn. The next message is planned again from scratch.

Availability

Starter Starter plan or above, the same as every MAS pattern.

What’s in this section

1. Editor tour — find your way around.
2. Your team — add the agents the planner may use.
3. Planning settings — models, plan size, and what runs at once.
4. Loops, failures & approvals — what happens when work comes back or a step fails.