Operator — an autonomous run with a ceiling
Open Operator →Operator takes a goal and works on it by itself: it plans tasks, calls tools, and writes one Markdown deliverable per run. The number that makes it safe to leave alone is the budget ceiling — every model call reserves against the run's remaining credits and is checked before the provider is called, so a run that would exceed the ceiling ends with the status budget instead of quietly spending more.
What Operator makes
One run produces one written deliverable, stored on the run itself. On request that same document is re-derived into named files and a .zip — so the artefact is the writing, and the file layout is a view of it rather than a second thing that can drift.
Runs happen on the server. Closing the tab does not stop a run, and reopening Operator re-attaches to it.
Walkthrough: one run, start to finish
Write the Goal
Up to 4,000 characters, and the single thing most worth spending time on. If you would rather start from a shape, the three examples — ResearchOp, PlanOp and BuildOp — fill the field with a goal that is already well-formed and you edit from there.
Attach context files, if the goal needs them
Up to five files, 20,000 characters each and 60,000 in total. These are what the run reads instead of guessing.
Choose a model
The default is Auto (first available), which is usually the right answer; pick a specific model when you care about a particular one's behaviour.
Decide on Fusion
Off, or 2, 3 or 4 drafts fanned out and then fused into one answer. Fusion buys diversity at a multiple of the cost, so it is worth it on a judgement call and wasteful on a mechanical one.
Pick the tools
reason is always on. Add search, image, code, data and graph as the goal needs. A tool the run cannot use is a tool it cannot waste credits on.
Set the budget ceiling and max tasks
Ceilings are 5,000, 25,000, 100,000 or 250,000 credits. Max tasks is 1 to 30, default 10. The ceiling is the hard stop; max tasks is how much structure you are willing to pay for.
Press Deploy run
Watch it work. Pause, Resume and Stop are live throughout. When it settles you get Continue, Run again, Share and Download .zip.
Options
- Model — Auto (first available), or a named model.
- Fusion fan-out — Off, 2, 3 or 4 drafts, fused into one.
- Tools — reason (always on), search, image, code, data, graph.
- Budget ceiling — 5,000 / 25,000 / 100,000 / 250,000 credits.
- Max tasks — 1 to 30, default 10.
- Context files — up to 5.
Limits
- One live run per account. Finish or stop the current one before deploying another.
- 4,000 characters for the goal.
- 5 context files, 20,000 characters each, 60,000 in total.
- 30 tasks is the ceiling on run structure.
- A shared run link exposes the title, goal, status, result and finish time — not the transcript and not what it spent.
- Only a settled run can be shared. A run still working has nothing final to show.
What it costs
Operator is a Pro app, and its spending is metered per call rather than estimated up front. Before each model call the run reserves the credits and checks them against what is left of the ceiling; if the reservation does not fit, the provider is never called and the run ends with status budget.
That ordering is the whole safety property. A ceiling checked after the call would be a report; checked before it, it is a limit.
Frequently asked questions
Can I close the tab while a run is working?
Yes. Runs execute server-side; closing the tab or losing the connection does not stop one, and reopening Operator re-attaches to it.
What happens when a run hits the ceiling?
It stops and settles with the status budget, keeping whatever it had written. Nothing is spent past the ceiling, because the check happens before the provider call rather than after it.
Is Fusion worth the extra credits?
On a question with one right answer, no — you pay two to four times for the same result. On a judgement call, a comparison or a piece of writing where the first draft is rarely the best, the fused answer is noticeably stronger.
What does a shared run show?
Title, goal, status, result and the time it finished. Not the step-by-step transcript and not the credits it used, so a link is safe to hand to someone who should read the output but not audit the account.
Can I run two at once?
No — one live run per account. It keeps the ceiling meaningful and keeps a runaway goal from competing with itself for your balance.
How do I get the output as files?
Download .zip. The deliverable is re-derived into named files on request, so the zip always matches the document rather than a stale copy of it.