Team Orchestration Patterns

A team's Team Type decides who works next and when a run ends. This guide covers what each type does, the settings that affect it, and tips for getting good results. Every field is listed in Team Settings.

In all six patterns, member order matters, and every member is described to the others by its team Role or, if that's empty, by its Default Role.

Sequential

Each member runs once, in member order. Each one sees the conversation so far, including earlier members' output.

You → Agent 1 → Agent 2 → Agent 3 → result

Use it for pipelines where each step builds on the last, such as plan → research → write → review.

Settings

• Member order is the run order. New members are added to the end.
• To reorder, remove members and add them back in the order you want. Sequential teams don't have drag-reorder.
• The form asks you to fill in routing fields (User →, Agent →) before you can create the team. They're pre-filled for you and don't change the run order.

Tips

• Put the agent that frames the problem first.
• Keep chains short, around 2 to 5 members. Every member adds to the context that later members read.

Concurrent

Every member receives the same input at the same time. Their replies are combined once all have finished. A team with a single member runs it on its own, as Sequential would.

        ┌→ Agent 1 ─┐
You ────┼→ Agent 2 ─┼──→ combined replies
        └→ Agent 3 ─┘

Use it for independent views of the same question, such as a legal, a financial and a technical review of one proposal.

Settings

• Drag members to set the order their replies appear in.

Tips

• Give each member a clearly different angle in its system message, so the replies don't overlap.

Group Chat

Members take turns in member order, round-robin, each building on the conversation.

Agent 1 → Agent 2 → Agent 3 → Agent 1 → ...

Use it for discussion, critique and iterative refinement.

How long it runs

Situation	Turns per message
Max Rounds set	Max Rounds turns
No Max Rounds, team waits for you (the default)	One pass: each member speaks once
No Max Rounds, team doesn't wait for you (for example Auto Respond, or Pause Between Turns on)	Two passes

Settings

• Drag members to set the speaking order.
• Leave Tool Agent Extra Turns off. It's a legacy turn order in which, after the first two members, only members with tools take turns.

Tips

• Three to five members with distinct perspectives works well, for example a researcher, a critic and a synthesiser.
• Set Max Rounds if you want several passes per message.

Group Chat Selector

A coordinator picks the next speaker each turn, or decides the round is finished.

coordinator → Agent 2 → coordinator → Agent 1 → coordinator → done

How the coordinator decides

• The coordinator is a model call on the first member's LLM configuration.
• It sees each member's name and role (the team Role, else the agent's Default Role, trimmed to 300 characters).
• It also reads the most recent messages, as many as Selector Max History Messages (10 when empty).
• It returns either the next speaker or a decision to stop.

How long it runs

• Until the coordinator ends the round, or until Max Rounds (three turns per member when empty) is reached.
• If the coordinator can't decide, for example because it named an unknown member, the team falls back to round-robin for that turn.

Use it for mixed expertise where you can't predict the order of work.

Tips

• Write a specific Role for every member. Roles are what the coordinator routes on.
• The first member's model does the routing. Make sure it's a capable model.
• The Selector type has no drag-reorder. To change who is first, remove members and add them back.

note

When a Group Chat Selector team runs as a Team node inside a workflow, it currently takes turns round-robin instead of using the coordinator.

Handoff

One member works at a time. A member passes control to another by calling a hand-off tool that the platform provides for this pattern. The first member starts.

Agent 1 ──(hand off)──→ Agent 3 ──(hand off)──→ Agent 2

When a member replies without handing off, the team waits for your next message. In Autonomous mode, or with Auto Respond selected, the team answers automatically instead, up to Max Autonomous Responses times.

Use it for triage, routing and escalation, where the current specialist knows who should handle the next step.

Settings

• Drag members to choose who starts.
• By default every member can hand off to every other. To restrict that, or to start with a different member, set handoff_rules or coordinator_agent_id through the API.
• GitHub Copilot agents can't be members of a Handoff team.

Tips

• In each member's system message, say when to hand off and to whom.

Magentic

The first member is the lead (manager), and the other members are workers. The lead:

1. makes a plan;
2. assigns each step to a worker;
3. tracks facts and progress in a ledger;
4. replans when the work stalls.

Lead ──→ Worker A ──→ Worker B ──→ Lead checks progress
  ↑                                       │
  └──────────── replan if stalled ────────┘

Use it for open-ended research and multi-step analysis, where the path isn't known up front.

Settings

Setting	Effect
Member order	The first member is the lead. Drag to change it.
Max Stall Count	Rounds without progress before the lead replans. 5 when empty.
Max Reset Count	How many full resets are allowed. Each one wipes the workers' memory of the task. 3 when empty.
Max Round Count	Total coordination rounds. Unlimited when empty.
Auto-approve Plans	Applies only when plan review is turned on through the API (enable_plan_review). When plan review is on and this is off, you approve the plan by replying "approve", "yes", "ok" or with an empty message, or reply with changes to ask for a revision.

If a Magentic team has only one member, the lead works alone.

Tips

• Give the lead a system message about the overall objective. Keep the workers narrow.
• Set Max Round Count to cap cost. Magentic uses more model calls than the other patterns.

Choosing a pattern

Scenario	Team Type
Fixed pipeline: research → write → review	Sequential
Several independent opinions on one input	Concurrent
Agents discuss and refine together	Group Chat
Route each turn to the right expert	Group Chat Selector
Triage and escalation	Handoff
Open-ended task needing a plan and recovery	Magentic

Human involvement in every pattern

Interaction Mode (Supervised, Autonomous, Collaborative) and its override fields work the same way in all six patterns. They control when the team waits for you, pauses between turns or answers on your behalf. They don't change tool approvals; see Teams and Using Approval Gates.

Related

• Teams
• Team Settings
• Building a Research Team