Agent teams
Let the AI split a large request into a team of focused sub-agents that work in parallel, then combine their results.
With agent teams on, the AI can assemble a small team for a big job. Each agent gets its own task, and the team's progress shows live in the chat.
What you can do
- Turn agent teams on or off for a project.
- Watch each agent's task, activity and summary in the chat.
- Stop the whole team or a single agent.
How a team works
For each job, the AI picks the roles it needs:
| Role | Access | What it does |
|---|---|---|
| Plan | Read only | Reads the project and lays out the steps before anything changes. |
| Research | Read only | Searches the project for the files and facts the other agents need. |
| Build | Edits files | Edits only the files it is given, so two agents never write the same file. |
| Verify | Read only | Checks the result and reports what still needs work. |
Up to 5 agents work in parallel. Each one reports back when it is done. Agents can share a task list, and a task can wait on another agent.
Turn agent teams on
Open the setting
Open Settings > AI > Agent teams, or use the Agent teams switch in the model picker.
Turn it on
Switch on Enable agent teams. The setting applies to this project. Agent teams are off by default.
Ask for a large job
Describe a multi-part feature, or a research, build and verify task. The AI decides when a team helps.
Follow and stop a team
The Agent team widget shows the team status (Working, Completed, Partly done, Stopped, Budget reached or Failed), each agent's phase, and how many agents and tasks are done. Open an agent to see its Task, Activity and Summary.
- Choose Stop team to stop all agents.
- Use the stop button on one agent to stop only that agent.
Ultracode
On Claude Opus 5.5 with a Claude Pro/Max sign-in, the Ultracode effort level turns on agent teams for the current session only. See Effort.
Teams have limits
A team has a token budget and a time limit. If it reaches one, it wraps up early and reports what it finished.