Skip to content

Agent Team Node ​

One agent can't be an expert in everything. Acme's front desk gets booking questions and billing questions, and each needs different knowledge, tools and tone. The Agent Team node lets several AI Agents work together on each message: a lead agent coordinates, and specialist agents do the work.

You'll find it under Core in the node picker, next to AI Agent.

An Agent Team with a Front desk supervisor, its AI Model, and two worker AI Agents (Bookings and Billing), each with its own Personality and AI Model.

How a team is built ​

The Agent Team node has four ports along its bottom edge:

PortWhat to connect
Supervisor *A Personality for the lead agent: who it is and how it should route or combine work. Required.
AI Model *The lead agent's model. Required.
WorkersOne AI Agent node per specialist. Click + on the port and pick AI Agent.
IntegrationIntegrations for the lead agent itself. These are used in the Refine loop and Parallel fan-out modes, where every agent in the team can call its own tools. In Team mode the lead agent only routes, so tools attached here aren't offered to it — give them to the workers instead.

Each worker is a full AI Agent with its own Personality and AI Model, plus any Knowledge, Integration and Action sub-nodes it needs. Workers can use different models from each other and from the lead agent.

Coordination Mode ​

Open the node to choose how the team works together:

  • Team (supervisor + workers) — the lead agent hands each message to exactly one worker, then answers with that worker's result.
  • Refine loop (generator + critic) — the Agent Team node drafts the answer and the first worker reviews it, round after round, until the critic is satisfied or Max Steps is reached. Any further workers are ignored.
  • Parallel fan-out (+ synthesizer) — every worker answers at the same time and the Agent Team node combines their answers into one reply.

The Agent Team panel with Coordination Mode set to Team and Max Steps 12.

Options ​

FieldWhat it doesOptionsDefaultShown when
Orchestration
Coordination Modeteam = supervisor+workers · refine = generator↔critic loop · parallel = concurrent workers + synthesizerTeam (supervisor + workers) · Refine loop (generator + critic) · Parallel fan-out (+ synthesizer)
Max StepsSupervisor agentic-loop iteration cap (default 12)

Max Steps counts differently per mode: supervisor iterations in Team mode, review rounds in Refine loop. The other settings in the panel (Worker agents, Routing Strategy, Enabled Modes and Parent Orchestrator ID) can stay at their defaults. Connect workers on the canvas through the Workers port.

What the team produces ​

The team's final reply is the agent reply for the turn, stored as {{ ai_response }}, just like a single AI Agent's. On Web Chat and voice it goes straight back to the customer. On WhatsApp, SMS and email, wire a sender for that channel so the reply is delivered.

Every worker's work is recorded in the run, so the run viewer shows which specialist handled each message.

When one of your people takes over a text conversation and then hands it back to the AI, the team picks it up again: it receives what the customer said while the person was handling it, so it can answer those messages without asking again.

If a team member fails mid-turn (for example a tool or the model returns an error), what the team had already done is still kept: its tool calls and the credits it used are recorded, and the conversation shows that the turn failed — instead of just the customer's message with nothing after it.

Testing a team, and runs with no customer message ​

A team answers text. When a run has no customer message, the team takes its task from the trigger instead:

  • a scheduled or event-driven run that carries an instruction for the agent: that instruction;
  • a Manual test run: the message or text field of the trigger's Sample Test Payload — for example {"message": "Where is my order?"}.

With neither, the run stops before calling any AI model and the run viewer says what to add. A customer message that is only an image or a voice note with no caption stops the same way, with an error saying the team answers text only.

Worked example: Acme's front desk ​

Setup. Asha adds an Agent Team after her Web Chat trigger. Its Supervisor Personality is Front desk: routes each request to the right specialist. She connects two workers:

  • Bookings: Personality Books and reschedules appointments, with Acme's booking integration.
  • Billing: Personality Answers invoice and refund questions, with the billing integration and the Billing FAQ Knowledge folder.

She leaves Coordination Mode on Team.

Action. Priya writes: "I was charged twice for last week's test."

Result. The Front desk hands the message to Billing. Billing looks up the payments and explains the duplicate charge. The Front desk replies to Priya with that answer.

What just happened. Priya talked to one assistant. Behind it, the right specialist, with the right tools and knowledge, did the work. Adding a new area later means connecting one more worker.

Agent Team or Steps? ​

  • Use an Agent Team when different kinds of request need different specialists.
  • Use Steps when one agent has to follow a fixed order, such as choose, check, collect, then book.

Next steps ​