Scheduled AI Tasks That Refuse to Start Work They Cannot Finish
Recurring agent work is easy to start and hard to trust.
The difference is what happens when the plan and the remaining headroom do not match.
Short answer
Give a schedule a brief and a subject rather than a cron line. The subject decides where the results land: on the run itself, on a task card that is re-run, or on a design. And a run that cannot finish inside its budget should defer, because a half-applied scheduled change is worse than none.
What this covers
- A schedule stands on one of three subjects and the subject decides where results go: on its own, on a task card that is re-run, or on a design whose plane each run lands on.
- Every scheduled run streams into a real conversation you can open, read and continue, rather than a log line.
- Before a run starts, a quota gate compares the headroom left against the reset time and what the task has historically needed; if it clearly will not fit, the run is deferred to the reset and you are told.
- Deferring is preferred over silently dropping to a weaker engine to finish, because a truncated result from a strong engine and a complete result from a weak one are both worse than running later.
- A task that has never run before is not deferred: it runs untimed the first time, because the honest thing to do with an unknown is measure it.
- Schedules exist as their own surface in Studio, in the browser and in the iOS app.
Comparison
| The situation | Finish at any cost | Defer to the reset |
|---|---|---|
| Thirty minutes of headroom, two hours of work | A truncated result you cannot use | A complete result, later, and you were told |
| Which engine ran it | Whichever one was still cheap enough | The one you chose |
| What you find in the morning | Something that looks finished and is not | Either the work, or a clear reason it waited |
Recurring work is easy to start
Ask for something to run every weekday morning and you have a schedule. The hard part arrives later, on the morning you open the app and cannot tell whether last night produced a real result or a half-finished one that stopped when a limit was hit.
Agents run against quotas that reset on a clock. That is the constraint every recurring agent job has to respect, and most of them do not.
Three subjects, and why that decides everything
A schedule stands on one of three things. On its own, as a standing brief whose results collect in its own history. On a task card, which is re-run on the repetition with every result collected on that card, so a recurring check becomes part of the board rather than a separate stream you have to remember to read. Or on a design, where each run lands on that design's plane and the design shows that a schedule is feeding it.
The subject is not a detail. It is what decides whether the output is something you will actually see.
The gate
Before a run starts, a quota gate compares the headroom you have left against the reset time and what this task has historically needed. If it clearly will not fit, the run is deferred to the reset and you are told.
The alternative most systems choose is to quietly drop to a weaker engine so the job completes. That produces something, and the something is not what you asked for. A whole-task decision should prefer a clean deferral to a silent downgrade, because a result you cannot trust costs more than a result that arrived later.
A task that has never run before is not deferred. It runs untimed the first time and becomes its own estimate, because the honest thing to do with an unknown is measure it rather than guess at it.
Every run is a conversation
A scheduled run is not a log line. It streams into a real conversation you can open, read and continue, exactly like a session you started yourself. That is the difference between a job that ran and work you can inspect.
And because Schedules exist in Studio, in the browser and on the phone, the moment you most want to check whether last night produced anything is not restricted to the moment you are at your desk. The [schedules documentation](/docs/studio/schedules) covers the subjects, the gate and the surfaces.