Teams and specialists

A specialist is one agent for one kind of work, known by its name, that you chat to, hand tasks to or add to a team. A team is the agents and people who take on work in a repository or a project, usually with an operations manager to plan it and check it. You build both under Teams.

The edit screens open with the same guidance as this page (How teams work, How specialists work), and the new team wizard has a short note at each choice with Learn more.

Make a specialist or a teamCopy Anchor Link

1. Open Teams and choose New specialist or New team, or start from a template such as Start with a pair or Feature squad.
2. Give each member a role, a runtime and a model, and say where it runs.
3. Write its instructions, or let a role fill them in.
4. For a team, choose its operations manager and how it reviews work; for a specialist, choose what happens to its finished work.
5. Save. You can change any of it later on the team's or specialist's page.

To use a specialist or team by default, pin it in the picker of Chat, Research or Team Work: new conversations there start with it.

People and agentsCopy Anchor Link

Each member is either an agent, which the app starts and gives work to, or a person, who does their part themselves. Agents and people share one board: either can be given a task, and the board shows who has what.

A specialist you add to a team joins as a link, not a copy, so a change to the specialist reaches every team that uses it. In a team, the operations manager gives it work.

What an agent runs onCopy Anchor Link

An agent has three separate choices:

• Runtime: the program that does the work, such as Claude Code, Codex, GitHub Copilot, Google Antigravity or OpenCode, or the built-in agent that comes with the app. See Runtimes and agents.
• Model: the runtime's own default, a hosted model, or a local model served on this computer by the bundled runner, Ollama or LM Studio.
• Where it runs: this computer, a node or another workbench in your organisation, or a provider's cloud. In a provider's cloud the agent works on its own and its work comes back as a branch for you to review. The cloud is offered only when you are signed in to Mojo Up AI Cloud or to your organisation.

Who pays for the workCopy Anchor Link

• Your own plan. A runtime signed in with your own account, such as a Claude, ChatGPT or Copilot subscription, uses your plan. There is no key to keep, and the work counts against that plan's limits. See Your subscriptions.
• An API key. Choose a provider for the member instead and it uses an API key kept under Settings → Models; that provider bills you for what it uses. A named key lets one team use a different account from another.
• A local model costs nothing per use.
• A node uses its own sign-ins, models and keys.

When a member reaches its plan's usage limit, it waits for the limit to lift, or hands its work to another account or runtime if you said it may. It is never shown as an error.

People on a teamCopy Anchor Link

A person on a team answers the questions agents put to them, does the tasks given to them, and approves what needs their say, such as an agent handing one of their tasks to someone else.

No agent runs for a person, so a task given to them waits for them: it is not stalled, nothing restarts it, and no agent takes it over. They find it in their Inbox with Start, Mark done and Blocked. If the team sets a time to wait, its operations manager is then asked what to do, but it never takes the task.

In a project your organisation shares, pick the person from your organisation's people. Otherwise name their GitHub or Azure DevOps account when the task board is kept in step with one.

Instructions and rolesCopy Anchor Link

A member's role says in a few words what it is for, and its instructions say how it works. A role from the built-in list or your organisation's fills them in, and a template's members come with theirs.

Every agent also reads the repository's own instructions for agents, and the team guidelines are given to every member when it starts. Keep what they all share in the guidelines rather than in each member. The skills you give a member add know-how for particular kinds of work, on top of its instructions. See Skills and the catalogue.

Good instructions are short and specific. Say:

• what the member owns;
• what done looks like;
• what it must not touch;
• when to ask.

The operations managerCopy Anchor Link

The operations manager is the agent that turns a request into a plan of tasks, hands each task to the member best suited to it, and reviews what comes back, approving it or sending it back with what to fix. Well-specified tasks can go to cheaper or local models.

A task still stuck after the review rounds you set comes back with the manager's proposal for what to do next: another approach, a split, accepting it with listed gaps, a change to the rest of the plan, or stopping. You decide in your Inbox, or let the team take the recommendation.

The operations manager is also the agent you talk to when you choose the team in Team Work. A team without one has you hand out the work.

A specialist has no manager. You choose what happens to its finished work: it merges, the specialist checks its own changes first, or each task waits in Review until you approve it. It can still ask you to look at a task when its own judgement says so.

Who answers whatCopy Anchor Link

When a team has people from your organisation, Who answers what on the team's page lists topics and the people who answer each, in order; who may cover for whom, always or while they are away; and how long is a reasonable wait.

Agents read these rules and choose whom to ask. The question goes to that person, and only they, or whoever covers for them, can answer it. Nothing moves to anyone else on a timer, and there is no fallback to the team's owner: after the wait, the agent decides whether to keep waiting, carry on or ask the next person, and says so.

Asking or decidingCopy Anchor Link

Agents decide what they can find out or work out for themselves, and never ask the same thing twice. A product, scope or design choice they must not make goes to a person as a question with options:

• with a recommended answer, the agent carries on with it and the person can overrule it later;
• without one, the task waits until the answer comes.

A tool call you set to Ask me pauses the agent until you answer.

Handing a task onCopy Anchor Link

An agent can hand a task to a teammate: the task moves to them on the board and their own session gets it with the context. Nothing new is started.

A person's task is never handed to an agent without their say: unless they marked it Blocked, they are asked first, in their Inbox.

When a task finishes, each task waiting on it gets a short note of what it did, the files it changed and what it left undone, so the next member does not have to guess.