Agent Teams

Your chat's model hands tasks to other models, then works from their reports.

The model you chose for the chat stays in charge. It can give a task to another model, or several at once, and carry on while they work. Each one starts fresh, does its one task, and sends back a short report. You choose who is on the team, so you decide how to balance cost, speed and load across the models you have.

Today, team members read and report. They look through files and documents, search and summarise, and the model in charge makes any changes itself.

Choose a team

1. Click the team icon in the chat header, or type /agents.
2. Choose New team….
3. Switch on the models you want. Move with ↑ and ↓, and press Space.
4. Press ⏎.

The team is named Team Blue, then Team Green and so on, and the icon takes its colour. It stays in the header's list, ready to pick on any chat. To give it your own name, press s in the panel, or choose Name.

A new chat starts with the last team you used. Choose No team to clear it for that chat, or × in the list to remove a team.

You can tap or click all of these too.

Use for

A task goes to a model at its level or a stronger one, never a weaker one. If two models could do it, your note helps decide.

Max at once

Max at once belongs to the model, so it stays the same in every team. It starts at 1 for a local model, 3 for a command-line assistant and 5 for a cloud model. Set it to what the machine or service can handle. Extra tasks wait their turn, and their row says so, for example queued (max 1).

Use it

Ask for it, or let the model decide when a job is worth splitting:

Use your team to read the three biggest files in this folder and summarise each one.

Each task appears as a row in the chat, with the model's name and a timer.

• Open a row to see that model's steps and report.
• ↩ sends your next message to that model. Choose Back to lead to talk to the model in charge again.
• Stop on a row stops that task. The main Stop stops them all.

You can keep typing while they work. The model in charge reads your message at its next step, without waiting for them to finish.

What teams are good for

Keep a long job's context small

Reading fills a model's context. Hand the reading to team members and the model in charge only takes in their reports, so it has room for the whole job. This suits a model with a long context doing the planning while others do the legwork.

Balance cost across your models

Give each kind of task to the model that suits it: a strong model for planning and hard questions, and local or low-cost models for lookups and summaries. Keep one team for each way of working, and pick the one that suits the job.

Give a small local model a bigger job

With no team, the model can hand a task to a fresh copy of itself. The copy works in its own context and the original waits, so this works even when your computer can only run one model at a time. No setup needed.

Use all your computers

A cluster shares the local models on your other computers with all your devices. Those models appear in the team panel under the computer's name, so one team can use the model on your PC and the model on your Mac together. Set max at once for what each computer can run.

Answer several questions at once

Tasks that don't depend on each other can run side by side, on separate models or computers. The more models on the team, and the higher their max at once, the more can run together.

Lead with a command-line assistant

Your computer (Circuit)

Choose a command-line assistant such as Claude Code as the chat's model, in the chat view, then pick a team. It can hand reading to your other models and get their reports back as a new message. Switching a chat to a command-line assistant turns its team off first, because these assistants have helpers of their own. Ask for your team when you want it used.

Command-line assistants can also be team members.

Good to know

• Your accounts and their terms. Each model runs on your own account, key or computer, under that service's terms and limits. Teams don't add to any limit.
• Limits. A model that reaches a limit waits and tries again, or reports that it failed. Its task doesn't move to another model on its own yet.
• Keep the page open. Team members run in the Circuitry page, so reloading it stops them. Their rows show as stopped.
• Terminal view. You can set up a team there, but the command-line assistant doesn't hear about it yet. Use the chat view.
• Phones with a Circuit. When a chat runs on your Circuit so that it can continue with the screen locked, it can't use a team.

Related

• Coding Agent
• Use your other computers: a cluster
• Using Local AI Models
• Use Your AI Subscription