Teams
A team is a group of agents plus an orchestration pattern, called the Team Type. The members decide what work gets done. The team type decides who speaks or works next and when the run ends.
You can chat with a team, use it as a Team node in a workflow, or test it like an agent.
Where teams live
Open Build → Teams. The Team Library on the left works like the Agent Library:
- + creates a team.
- Add folder creates a folder.
- Drag a team into a folder to move it.
Teams belong to the current project, and so do their members.
Team types
| Team Type | How it runs | Good for |
|---|---|---|
| Sequential | Each member runs once, in member order. Each sees the conversation so far. | Pipelines: plan → research → write |
| Concurrent | Every member gets the same input at the same time. Their replies are combined when all have finished. | Independent perspectives on one question |
| Group Chat | Members take turns in member order, round-robin, building on each other's messages. | Discussion, critique, iterative refinement |
| Group Chat Selector | A coordinator reads the conversation and each member's role, then picks who speaks next or ends the round. | Mixed expertise where the order isn't known up front |
| Handoff | One member works at a time and passes control to another through a tool call. The first member starts. | Triage and escalation |
| Magentic | The first member is the lead. It plans, assigns work to the other members, tracks progress and replans when things stall. | Open-ended research and multi-step analysis |
Team Orchestration Patterns explains exactly how each pattern behaves and which settings affect it.
Members and order
Add members with Assign Agents and remove them with the delete icon on their chip.
- Order matters for every team type. It sets the run order for Sequential and Group Chat, the start agent for Handoff, the lead for Magentic, and which member's model the Group Chat Selector coordinator uses.
- New members are added to the end.
- For Concurrent, Group Chat, Handoff and Magentic teams, drag the member cards to reorder them. The new order is saved straight away.
- Group Chat Selector teams show a Role field for each member instead. The coordinator reads it when choosing the next speaker. If it's empty, the agent's Default Role is used.
- Sequential teams show routing fields (User → and one per agent), which must be filled in before you can save. The pattern itself runs members in member order. To change a Sequential team's order, remove members and add them back in the order you want.
How people take part
Interaction Mode sets how much a person is involved while a team runs:
| Interaction Mode | What it means |
|---|---|
| Default (Supervised) / Supervised | When a member asks for input, the team waits for you. Group Chat teams do one pass through the members per message you send. |
| Autonomous | When a member asks for input, the team answers automatically, up to Max Autonomous Responses times (default 5) before it stops. Magentic plans are approved automatically. |
| Collaborative | The team pauses after every member's turn. You can type continue (or press send on an empty message) to carry on, type stop to end, or type anything else to steer the conversation. |
When you choose an explicit mode, more settings appear: Pause Between Turns, User Input Handling, Max Autonomous Responses and, for Magentic teams, Auto-approve Plans. Each one overrides the corresponding part of the mode.
User Input Timeout (ms) sets how long the team waits for a reply before it carries on. Waiting for a tool approval doesn't time out.
Interaction Mode doesn't change whether tool calls need approval. That's decided by approval pins and the risk appetite; see How approval is decided. In unattended runs, such as a workflow, a tool call that needs approval is refused rather than approved.
Other team settings
- Approval Override: pins tool-call approval for every member. A team pin takes precedence over agent and function pins, except that critical-risk calls still need approval.
- Enable Streaming: streams members' replies token by token. This replaces each agent's own Streaming setting.
- Tool Summary Max Characters: tool results longer than this are saved to the session and replaced with a pointer the model can page through. See Built-in Tools.
- Introduction Message: the team's opening message.
Every field is listed in Team Settings.
The team page
The header has the Active / Killed killswitch and Open in Chat. The tabs are:
| Tab | Contents |
|---|---|
| Overview | Summary cards (Members, Risk and compliance, Readiness, At a glance, Reach, Activity · 30 days, Tests), followed by the team's settings and members, which you edit here |
| Configure | Currently empty. Edit settings on Overview. |
| Topology | What the team's members can reach |
| Tests | Evaluation test cases; see Testing Agents and Teams |
| Observability | Usage, cost, errors and audit trail; see Observability |
| Risk | Risk classification and review |
| Controls | Governance control evidence; see Governance |
| History | Configuration revisions and Restore |
Copy Team duplicates a team's settings and members. Deleting a team requires the project Admin role.
Running a team
- Open in Chat starts a conversation with the team. See Chatting with Agents.
- Add a Team node to a workflow.
- Through the API, start a group chat session for the team (
POST /chat/startwithchat_type: "group"and the team's ID). Teams don't have a one-shot run endpoint like agents do. See Platform API.
A killed team can't start new sessions until it's reactivated with the Active / Killed switch. Killing it can also end its active chat sessions (Also terminate active chat sessions for this team).
Templates
Platform teams can appear in your library marked Managed by platform — read only. You can also add teams from the Marketplace. Platform administrators can publish a team as a template with Publish Template. See Templates.