A team is a named group of agents built around a purpose — "Research", "Blog Team", "Sales."
The coordinator delegates work to a team by name; the team then runs its own tool loop to complete
it and hands back a result.
Every team has:
- A name and description — what it's for. This is what the coordinator reads to decide whether a task fits.
- A status (healthy / idle / error / deploying) and a category — a workspace/group label used to organize teams.
- Permissions — the scopes this team's agents are allowed to use.
- One or more agents — the actual workers.
- Skills attached to it (see Skills).
- Its own Brain folder, created automatically at
Teams/<team-name>/for work logs, config, and outputs.
Teams can be user-created or seeded by the platform. A few teams are marked as system teams
(like the File Manager) — these exist for platform housekeeping and aren't meant to be deleted.
Each agent on a team has:
- A name and optional description.
- A model — which AI model it actually runs on (for example
claude-sonnet-4-6). By default an agent inherits the install's default model, but this can be pinned per agent. - modelRequired — whether that model is pinned. If you tell the coordinator "make the Researcher use Claude Opus," it sets
modelRequired: trueon that agent, and from then on it always runs on that model instead of falling back to the default. - Capabilities — a list of tags describing what it can do.
- An optional system prompt — its specific instructions/persona.
- Temperature and max tokens — standard generation controls, with sane defaults.
You do this by asking the coordinator in plain language — for example:
The coordinator has tools to create teams and agents, add skills, update an existing team's name or
status, update an agent's model or system prompt, and list a team's full details when you want to
review or change what's already there.
Agents aren't locked to one AI provider. You can mix and match freely in the same install — the
coordinator on Claude, one team on a free NVIDIA model, another agent on a local Ollama model — see
Providers & Models for the full picture.