# How to put an agent on a schedule

Add a routine so your agent runs by itself. Set a cadence like Daily 09:30, review what each run actually did, and approve anything that needs you.

Routines run an agent automatically, on a schedule or a trigger. The setup is short. The useful habit is what comes after: reading the runs.

1. Open the agent and add a routine.
2. Set the cadence in the schedule builder. Cadences look like "Mon 09:00" for a weekly run or "Daily 09:30" for every morning.
3. That's it. From here the agent runs without you opening GPTree.

Routines can also fire from a trigger instead of a fixed time, for work that should happen when something happens rather than at nine every day.

## Checking what a run did

Each routine shows its runs, so the routine itself is the first place to look. For the wider picture, the Activity feed and run views show what the agent actually did across everything, not just this one job.

Read the first few runs closely. A schedule means the agent works while you're not watching, and the fastest way to trust that arrangement is evidence, not hope. Start with a daily cadence, skim Tuesday's run on Tuesday, and only stop checking once the output has earned it.

## Saying what a routine reads

A routine can declare the sources each run has to read: a connected app such as Slack or Linear, the web, or your knowledge sources. When you ask an agent in chat to set up a routine, it fills this in from what the work needs, and the routine editor shows the list under Reads. You can remove an entry there before you save. The editor does not add entries.

The list does two things. A run that cannot reach one of its sources, because the app is not connected for example, is blocked before it starts: it costs next to nothing, its result says what is missing, and you get a notice. And a run that reaches its sources but reads none of them is marked "Not verified", so a report written from nothing does not pass for a report. See What "Not verified" means on an agent result.

One limit follows from where routines run. A routine runs in the cloud, so it cannot read files, a repository or a database on a computer. A routine that declares one is not saved, and an agent asked to draft one will tell you so instead of scheduling work that would run against nothing. If the same information is in an app you have connected, such as a code host or an issue tracker, have the routine read that app.

## When a run is waiting on you

A scheduled agent doesn't get extra permissions. Write actions, like sending an email or creating an issue, still stop and ask for your approval, and anything the agent needs from you (questions to answer, actions to approve) waits in the Action Plane, your inbox.

So if a routine seems stalled, check Needs you before assuming something broke. It may just be waiting on a yes from you, and the approval will have "What led to this" context attached so you can see how it got there. When nothing's pending, Needs you says "You're all caught up."

One planning note that follows from this: if a routine's whole job ends in a write action, its schedule really means "prepared and waiting by 09:30", not "sent by 09:30". The last step is still yours.

## Pausing one routine

Every routine has its own switch. You find it in the agent's Routines panel, and on the Work page under Agents, where each routine is a row showing its cadence, the agent that runs it and when it runs next. Turn the switch off to pause that routine alone; turn it back on to resume it. The rest of the agent's work carries on either way.

## When a routine doesn't run at all

Routines only run by themselves while autonomy is on. That means the Autonomy switch for your account is on, and so is the agent's own "Run tasks automatically" setting. While either is off, a scheduled or webhook routine still comes due, but the run is skipped and filed as a to-do instead of being worked. If you turn your switch on and it still reads Off, autonomy is switched off globally, which only an admin can change.

GPTree tells you when this is happening instead of letting the routine look healthy. In the Routines panel, any routine it affects shows "Won't run: autonomy is off, or Run tasks automatically is off for this agent" under its schedule, the next-run time is hidden (that run would be skipped), and the Autonomy switch appears right above the list so you can turn it on without leaving the agent. The first time a routine is skipped this way you also get one notice, in GPTree and by email, whatever that routine's Notifications checkboxes say. You won't get one for every missed run: the next notice only comes if the routine runs again and later gets skipped again.

Routines set to Ask every time are the exception. They never run unattended anyway, so autonomy doesn't hold them back; each run waits for your approval instead.

## When your agents are paused for spending

There is one more reason a routine can be held, and it is not a switch. GPTree puts an hourly and a daily ceiling on how many credits your agents may spend on work you did not ask for, and reaching either one pauses that work until you resume it.

The ceiling covers unattended work only: scheduled and webhook routines, and tasks an agent picks up by itself. Crews that start on a schedule or a webhook are not covered yet, so a crew keeps running when the ceiling pauses everything else. It never touches your own chats or anything you press Run on, which stay limited only by the credits on your plan. That distinction is the point of it. A single agent can start many separate chains of work at once, and each chain gets its own runaway limits, so a backlog waking up together could spend heavily in a few minutes with every individual limit behaving correctly.

You find out once, by email, naming the ceiling you reached. In the Routines panel a held routine says the run was skipped because of the credit ceiling, which is deliberately different wording from the autonomy notice: your switch is already on, so turning it on is not the fix. The Autonomy control shows Paused with a Resume button, and pressing it starts your agents again straight away.

Nothing is thrown away while you are paused. Each occurrence is still recorded as a to-do, so resuming picks the work back up on the next cycle. If it pauses again shortly after you resume, open Usage and Cost and look at what the agents were spending on before setting them going again, because a routine that costs more than you expect every time it fires is worth changing rather than restarting.
