Skip to main content

Agent Teams

An agent team (called an "app" in older docs) is a group of agents working together, with one of them designated as the entry point. You'll find them in the product under Build → Agent Teams.

A single agent can answer "what's our refund policy?". An app can answer "I want a refund for order #1234, my email is X" — because behind the scenes the entry-point agent might call a Lookup Order agent, then an Issue Refund agent, then a Send Confirmation agent.

The user sees one chat. Under the hood, three agents did three jobs.

Why apps exist​

You could build everything as one big agent with twenty tools. It works, until it doesn't:

  • Big agents are slow — they re-read all their instructions on every turn.
  • Big agents drift — too many concerns in one prompt and behaviour gets fuzzy.
  • Big agents are hard to change — touching the prompt risks breaking unrelated behaviour.

Apps split the work. Each agent has a clear job. The entry-point agent decides which specialist to call.

Creating an app​

Go to Build → Apps → New App.

You'll set:

  • Name and description
  • Entry-point agent — the one that receives the user's message first
  • Status — draft or published

That's it. The app exists.

Wiring specialist agents​

An app doesn't have a separate "specialist list". The entry-point agent has agent-tools attached to it (see Agents). Those agent-tools are the specialists.

Open the entry-point agent, go to Tools, and attach the agents you want available. The app shows the resulting topology — entry point at the top, specialists below.

Topology view​

Open any agent team and you see the Entry Point Agent at the top and an Agent Topology panel below — a tree showing who can call whom, auto-derived from the entry-point agent's tool configuration. Team Members lists every agent in the team.

Agent team page: Entry Point Agent picker, Agent Topology tree, Team Members list

Useful for:

  • Spotting unused agents
  • Spotting circular calls (Agent A calls B which calls A)
  • Onboarding new team members to how a multi-agent system fits together

Publishing​

Apps start as draft. Drafts are visible only to your tenant. Published apps can be:

  • Exposed via a public chat URL (see Sharing)
  • Called via the API
  • Hosted as MCP servers
  • Triggered via webhook (see Triggers)

Switch from draft to published when you're confident. Switch back to draft to take it offline without deleting it.

Agent Team Runs​

Every time someone uses an agent team, the run is recorded under Work → Agent Team Runs. You can see:

  • The full conversation
  • Every agent that was called
  • Every tool call inside each agent
  • How long it took
  • The final result and overall status (Running / Failed / Cancelled / Completed)

Agent Team Runs list: Run ID, team name, status, progress, trigger, started, duration

See Monitoring for what's there.

Editing live apps​

Like agents, apps are live. Changes take effect on the next run. The entry-point agent's instructions are the most sensitive — change those carefully.

Use clone to test changes safely.