Squads and Multi-Agent

Instead of a single agent trying to cover everything, you can build a squad: several specialized agents with an orchestrator that receives the conversation and transfers to the right specialist. This is the recommended pattern when the support scope grows.

When to use squad vs. single agent

Use single agent when:

• the scope is narrow (e.g., only support for one product);
• the prompt fits without becoming huge;
• there’s no need for specialization by area.

Use squad when:

• you need to serve different areas (sales, support, finance);
• a single agent’s prompt would become too long and contradictory;
• you want different models for different tasks (cheap for triage, strong for resolution);
• some conversations need to go to a human at specific points.

Orchestrator + specialists pattern

The most common setup:

1. Orchestrator agent — receives the conversation, identifies user intent, and transfers to the right specialist. Usually a cheap model with a short prompt focused on triage.
2. Specialist agents — each covers one area (e.g., Technical Support, Sales, Finance). Each has its own tools, knowledge bases, and tone.
3. Human handoff (optional) — any specialist can hand the conversation off to a person via webhook when the case is out of scope.

How transfers work

Agent-to-agent transfer is done by the native tool Transfer to agent.

1. Open the agent that should transfer (usually the orchestrator).
2. Go to Tools → + Add Tool.
3. Pick Transfer to agent.
4. Toggle the switch and check the target agents — only active agents from the same organization appear (and the agent being edited doesn’t appear, to prevent self-transfer).
5. Changes persist automatically.

Details in Transfer to Agent.

When the tool is called:

• the conversation changes owner — the new agent takes over and answers from that point;
• history is preserved and the new agent sees everything said before;
• transfer is one-way — to go back, the destination agent must have the tool too.

Human handoff

To hand the conversation to a real person, enable the native tool Human Handoff on the agent. It requires an HTTPS webhook URL — your organization receives a POST with the conversation context when the agent calls the tool.

After handoff, the agent stops responding in the conversation (ai_enabled = false) until someone reactivates manually from the Conversations panel or via API. Useful to integrate with CRM, Zendesk, Slack, or any tool your team already uses.

Details in Human Handoff.

Design tips

• Keep the orchestrator lean. It only needs to decide where to transfer, not solve the problem. Short prompt, cheap model.
• Name specialists clearly. The orchestrator uses the name and description of the target agent to decide — Sales SDR is better than Agent 2.
• Avoid cycles. If A transfers to B and B transfers back to A without criteria, you’ll loop. Model the flow as a tree, not a graph.
• Each specialist has its tools. Don’t give everything to everyone — the more focused, the better the agent decides.
• Test together. Use Test Agent on the orchestrator and simulate real conversations that need to go to each specialist. Verify the transfer happens when it should.
• Mind the cost. Each transfer adds an extra LLM call. At high volume, this adds up.