Agent teams

{
  "handoffs": [
    {
      "name": "Billing",
      "agent_id": "b41c7a29-6f2b-4c7a-9c1e-1d2a8f0b41c7",
      "description": "invoices, refunds, payment questions and card details",
      "context": "summary",
      "message": "Let me put you through to billing."
    }
  ]
}

​

When to split one agent into several

• Prompt scope. One prompt covering billing, returns and technical support makes the model hold three policies at once.
• Models or voices per desk. A fast, cheap model on the front desk and a stronger one where it reasons about a refund; a different voice or language.
• Tools per desk. The refund tool lives on the billing agent only, so the front desk cannot call it.

​

What crosses a handoff

• userdata — session state every agent on the call shares.
• vars — one bag of values per session, merged over each agent’s own declared defaults.
• Conversation history — as far as the edge’s context policy allows.
• Speech streams, when they match. A target with the same STT, language and endpointing keeps the listening stream, so a sentence in progress is not cut off; one with the same TTS model and settings speaks through the same connection. A target that differs opens its own.
• The call record. One call, one transcript, with an agent_handoff item at each switch and every item stamped with the agent that produced it.

​

Next