Auto Pilot & Accept All
Several Studio pages generate a list of artefacts rather than a single answer. Accept All takes the click-per-item tax out of reviewing one batch. Auto Pilot goes further: it runs the generation over and over, accepting as it goes, until the agent has nothing new to add.
Accept All
Wherever an agent proposes a list, an Accept all (N) button sits alongside the per-item Accept buttons. It is available in every operating mode — accepting a batch is still a human decision, just one made about several things at once. You will find it on the Ideas Board generate panel, on all three Shape steps, on Vision outcomes, on Ask Agent replies, and on Context Library suggestions.
Auto Pilot
Auto Pilot accepts AI output with no human decision on each item. That is precisely what Supervised mode exists to prevent, so the 🛸 Auto Pilot button only appears when the studio's Operating Mode is 🤝 Semi-Autonomous or ⚡ Fully Autonomous.
Each surface runs its own pipeline:
"Take it from here" — the end-to-end journey
Open an intent from the Intent Pad and a Take it from here banner offers to run the whole thing. From a single click Auto Pilot will:
- Write the intent's brief, if it does not have one.
- Break it into workstreams on the intent.
- Create one Work Package per scope under each workstream, one workstream at a time.
- Generate work items on each package's own page, one package at a time.
In a Semi-Autonomous studio the run ends there. Every package it made is left as a draft, including one that ended with no work items: sending a package for review, and starting its readiness gate, are yours to do on each package you want to take forward. Auto Pilot sends nothing for review and archives nothing itself.
In a Fully Autonomous studio the run goes on: the reviewer agents assess each package it made, and under the owner's standing approval it approves and dispatches those that pass and starts their coding runs. A package whose gate has not passed, that a route refuses, or that the planning check still questions is left for a person and listed on the run with the reason (see Checking the Codebase Before Planning).
It never moves you. The panel shows where it is working — "the work items on work package Send the link" — with a Go to it link that opens that page at that section, so you can watch the studio fill in when you want to and carry on with something else when you do not. The panel docks to the side of the screen and follows you across every page, which means Stop is always to hand; close it mid-run and a bar at the top of the studio says Auto Pilot is running with a View progress button. The run finishes with everything it created outlined and labelled on the Work Packages board — in a Semi-Autonomous studio each one a draft, ready for you to open, check and send for review.
The journey runs on the server, so it carries on with the tab closed and several intents can run at once: see Auto Pilot on the server below, including how a stopped run is picked up again.
The request limit — how many passes a run makes before it pauses to ask, each pass making at most one AI request — is a studio setting (Trust & Autonomy, owner-only, 10 to 1,000, 100 by default).
Running without the prompt
The panel carries a Keep going without asking checkbox. Tick it and the run-budget prompt never appears — useful when you would rather not sit and wait for it. You can tick it before starting or while a run is in flight; ticking it while the prompt is showing answers it and carries on. It suppresses the question only: item caps, the empty-pass rule and the error caps still end the run, and Stop stays live throughout.
The panel
Every page uses the same panel, so it behaves identically wherever you meet it: a progress bar and headline counters, a pipeline showing each stage and its state, a live feed of items as they are accepted, and Pause and Stop controls. Pause parks the loop between passes; Stop ends it, letting the in-flight pass finish so its items are kept rather than orphaned.
Each item in the feed carries a Remove button and an ℹ️ Details toggle that expands the artefact itself — a workstream's assumptions, a slice's acceptance criteria and risks, an outcome's key results, a context entry's text. Deciding whether to keep something is a judgement about its content, so the content is never more than one click away.
When the run ends the panel becomes a review list of everything it added, with Remove on each row and Remove all beneath. Removing an item deletes it from the store it was written to — this is a genuine undo, not a display filter.
How it knows when to stop
Repeated generation against the same material is a money pump if it never terminates, so four guardrails are live at once:
Why it does not repeat itself
An agent asked the same question twice will happily answer it the same way twice — which would both fill your studio with restatements and stop the loop ever terminating. Two things prevent it. Every generation sends the agent an explicit list of what already exists with an instruction not to re-propose any of it, and an empty result declared a valid answer. Then everything coming back is checked again on the client and near-duplicates are dropped, with the count shown so you know it happened.
Where the loop runs
On the Ideas Board, Vision and the Context Library, Auto Pilot runs in your browser. Closing the tab or navigating away ends the run — deliberately, so it cannot keep spending against your AI provider with nobody watching. Everything it accepted before that point is already saved.
The intent journey — Take it from here, Auto Pilot this workstream and Auto Pilot this package — runs on the server, described next.
Auto Pilot on the server
An intent's Auto Pilot is sequenced by the server, not by your browser. Send it and close the tab: the run carries on, and you come back to the intent with the work done or the run waiting for you. Several intents can run at once — one run per intent, up to three in a studio — and everyone in the studio sees each run, who started it and how far it has got.
What it may cost, before it is sent. A dialog shows the run's limits and costs:
Nothing is sent until you tick that the run is billed to the studio.
Checking the codebase first, and checking the plan. When a repository is connected the dialog also has a Check the codebase first box: a run on an intent or on one workstream then starts by reading the repository for what already exists, and plans the change to what is there. When the run plans the intent (or the workstream), a second AI request checks the plan against what you asked for before anything of it is written, and the planner revises it at most twice. Planning may then take up to six AI requests, all counted in the limit above. Both are described in Checking the Codebase Before Planning.
While it runs. A bar at the top of the studio and the Activity panel list every run, with its stage, passes and AI spend; View progress opens its panel, which shows the stages, what the run has made and the Workbench credits it has used. The person who started a run, the studio's owner or a reviewer can stop it. When a run reaches its limit it waits for the person who started it to let it carry on for one more step (up to ten times) or stop it; a question left unanswered for a week stops the run. No browser needs to be open while a run goes.
Picking a run up again. If a run is stopped, or the service cuts it short, the person who started it can Resume it: it carries on from its first unfinished stage as the same run, and what it had already made is kept. A run an owner or reviewer stopped stays stopped for a contributor who started it; a run can be resumed up to ten times; and a run that stopped on a finding — the intent changed under it, or an answer failed its quality check — is a new run's job.
The review list. Remove takes a workstream off the intent or a work item off its package, and archives a work package rather than deleting it. A work package the run left with no work items stays a draft: a run archives nothing itself. A finished run's record is kept for 90 days; what it made is yours for good.