Autopilot — setting it up
AI project management, configured per project, in five steps.
Autopilot does the project-management work nobody enjoys: turning a brief into a real plan, routing work to whoever actually has capacity, chasing status, and keeping the schedule honest. It runs per project.
Autopilot is configured on the Noot it runs on, and nowhere else. Open the project and use the Autopilot settings button in its header (it reads Plan with Autopilot when it's off), or its Autopilot tab. Anyone looking for it in Settings or Org settings won't find it, because a single switch for every project in a company was never the right shape.
You can configure it if you have the AI configuration permission or you created the Noot. It's on the Business plan and above.
The five steps
| Step | The question | What you set |
|---|---|---|
| 1. Brief | What are we delivering, by when? | The strategic brief, delivery date, budget, project type and methodology. |
| 2. Teams | Who's on it, and what do they bring? | A project-local specialty for each team — what they're doing *here*. |
| 3. Autonomy | How much may Autopilot do alone — and how far ahead? | The autonomy level, the horizon, twelve capability switches and a cadence. |
| 4. Guardrails | What must it never do? | VIPs, stakeholders, communication hours and scheduling limits. |
| 5. Access | What may it reach? | Calendar, GitHub and chat grants. |
1. The brief is the whole game
A weak brief ("Update the pricing page") produces a weak plan. A strong one names the outcome, the deadline, the constraint and what done means. The delivery date is a real constraint — the plan is compressed to fit inside it, so a date nobody meant is the single most expensive thing you can get wrong here.
You don't have to write it cold. Help me write this runs a short interview — five questions, one at a time: what are we delivering, by when, what's the hard constraint, what does done look like, and what should it watch out for. Your answers become a real brief you can then edit.
The AI writes the prose; the dates, budget and project type are read by the app, not by the model. Where the app can't read what you wrote — "end of Q3" — and the model resolves it instead, you're told in words what it read, next to an editable field. Nothing is quietly decided for you.
Focus the brief box and a hint opens showing a weak brief and a strong one side by side, with one click to start from the strong example. It stays folded until you ask, so it never sits between you and the field.
3. How far ahead, and how much on its own
There are two independent axes here, and it's worth being clear which is which.
The horizon — how much of the project it answers for
| Horizon | Behaviour |
|---|---|
| Up to a milestone | It plans and manages the work leading to one milestone, then stops. Anything due after that milestone is left alone — not planned, not graded, not re-dated, not reassigned, not chased. This is the default. |
| Start to finish | It runs the whole project out to its delivery date. |
It has to plan that far ahead, and it can only plan from what you tell it — so it asks for a sentence on every open milestone before you can pick it, and that detail goes into the plan rather than sitting in a box. Every project configured before the horizon existed reads as up to a milestone, because widening what an autonomous agent may touch is not something a change of ours should do to you.
Undated work stays in scope either way: a task with no date isn't evidence that it belongs to a later phase.
The autonomy level — how much it may do without asking
| Level | Behaviour |
|---|---|
| Draft & Propose | Builds the board, picks owners, queues invitations — then stops. Nothing applies and nobody is contacted until you approve it, item by item. This is the default. |
| Draft, and apply safe changes | Tasks from explicit meeting action items, priority bumps, check-ins and risk logs apply on their own. Reassignments, date moves and meetings still wait for you. |
| Fully Autonomous | Creates tasks, messages people for updates and shifts timelines by itself. Every change is still listed, and individually reversible, in the activity trail. |
Underneath, twelve independent switches decide what it may do at all — create tasks, assign members, build boards, set due dates, reprioritize, schedule meetings, manage goals, ask owners for status, follow dependencies, shape this project's teams, log risks to RAID, send recaps. Turning one off genuinely stops the behaviour, it doesn't just hide a button. Two are off by default: *Manage goals* and *Send recaps*.
*Shape this project's teams* lets Autopilot name a lead for a team that has none, or group un-teamed people into a project team when your organization's teams don't cover the work. Both are proposals whatever your autonomy level says, including Fully Autonomous — a team is an organizational object, and creating one behind somebody's back is not a change an agent should be able to make on its own.
Cadence is how often it reviews the project by itself: Manual (the default — you run every cycle), Daily or Weekly. Cadence is per project, so a daily Noot and a weekly Noot in the same workspace genuinely behave differently.
4. Guardrails
- VIPs — people Autopilot may never hand heavy work to: a CEO, an external client, anyone whose time isn't yours to book. You set the heaviest task a VIP may be given, in hours; at zero they're never auto-assigned anything. This is enforced when work is assigned, not merely requested of the model.
- Stakeholders — they get a weekly executive summary of the project whether or not they own a single task.
- Communication hours — when Autopilot may put a message in front of a person, read in each recipient's own timezone. Outside the window a message waits rather than being cancelled; it goes out on the next cycle inside the window, and a person can always approve it by hand.
- Scheduling limits — respect working hours, respect time zones, no weekend meetings, and require approval for meetings. Leave that last one on and nothing is booked without you, at any autonomy level including Fully Autonomous; a new project starts with it off, so Autopilot books the rhythm itself. There's also a cap on how many tasks a single run may create.
5. Access
Three grants, each unlocking one specific behaviour. Revoke one and that behaviour stops — the data isn't merely hidden, it is never read.
| Grant | Unlocks | Default |
|---|---|---|
| Calendar | Booking and running the project's meeting rhythm — see What a cycle can do. Revoke it and no meeting is booked. | On |
| GitHub | Reading merged pull requests linked to this project's tasks. | Off |
| Chat | Reading recent messages in conversations pinned to this Noot, for blocker context. | Off |
Noots has no design-tool integration, so a switch offering one would do nothing. The screen says so rather than listing a grant that isn't real.
The Access step reads live connection state, so if you enable a grant with nothing connected it tells you on the spot. A workspace with nothing connected is not a broken one: Autopilot makes a Noots room for the meeting and sends the in-house bot to it, which is the floor rather than a failure state.
If your workspace runs its meetings on Zoom, Google Meet or Microsoft Teams, Autopilot can create the meeting through the organiser's own connected account — which means it appears on their calendar, as them. That needs a person to agree to it: there is a checkbox on this project's Autopilot settings, it is off until somebody ticks it, and it is re-checked on the server before any token is looked at. Declining costs you nothing except the platform: the meeting is still booked, still has an agenda and still gets a bot, in a Noots room.