Schedules
A schedule is a clock pointed at an agent or a workflow. It adds no intelligence of its own — it just means you stop being the trigger. Weekly pacing reviews, morning digests, anomaly sweeps every six hours: anything you’d otherwise have to remember.
In one line
Pick who runs it, tell them what to do, say when. The output lands as proposal cards, a document, or a Slack message — never as a silent change to your accounts.
Set one up#
From AI Studio → Schedules → New Schedule. There are four things to fill in.
Who runs it#
Choose an Agent or a Workflow as the target. Agents suit jobs where the situation varies and judgment matters; workflows suit fixed pipelines. If you’re not sure which you need, see when to use what.
What to do#
For an agent, write the instruction it should receive on every run — the same thing you’d type into a chat. Type / to pull in a skill, and @ to name a specific ad account. Files you attach are re-sent on every run.
A skill picked with / is loaded on every run, not merely suggested — it is how you make the run follow a fixed process. Only skills the chosen agent has are offered.
Write it as a standing instruction, not a one-off: “Review this week’s budget pacing across all campaigns. Identify under- and overspending. Propose reallocations to maximise ROAS.”
For a workflow, you supply its Input instead — the JSON payload the trigger expects.
When#
Set the cadence — every N hours or days, specific weekdays, or a day of the month — plus the time and the timezone it should be read in.
Save it#
The schedule appears in the list with its next run time. You can pause and resume it at any point, or trigger a run by hand to check it works before waiting for the clock.
Starting from a preset#
The Schedules page ships with presets for the jobs most teams set up first — a daily performance digest, a weekly budget review, an anomaly alert every six hours, a monthly executive report. Each one arrives with a cadence and an instruction already written, which you can edit. They’re the fastest way to see the shape of a good scheduled instruction.
What a run can and can’t do#
Scheduled runs propose; humans approve
A scheduled run has no human attached, so it can’t push a change to Google, Meta, or Microsoft directly — every real platform change goes through an approval preview, and there’d be nobody there to answer it.
What it can do is write proposal cards into your decision queue, and to Slack if you’ve connected it, with the reasoning and estimated impact attached. Give the agent the Suggestions capability and brief it to propose, not to “fix” — an instruction like “pause the losers every morning” will correctly turn into cards you approve at 9am.
Runs that produce a document work the same way: with the Artifacts capability on, the PDF becomes a real file the run can hand onward — emailed, posted to Slack, or waiting on the thread when you open it.
Watching it work#
Open any schedule to see its run history: when each run fired, whether it succeeded, and how long it took. Open a single run to read the whole conversation it had — what it looked at, what it concluded, what it proposed. When a scheduled job starts producing output you don’t like, this is where you find out why, and the fix is almost always a sharper instruction.
Common flows#
- Monday pacing review. A Budget Analyst agent, weekly, proposing reallocations into your decision queue.
- Morning digest. A reporting agent, daily at 8am, posting yesterday’s numbers to a Slack channel.
- Anomaly watch. Every six hours, flagging any metric more than 20% off its 7-day average, ranked by budget impact.
- Client pack. A workflow, monthly, that writes the analysis, renders the PDF, and emails it to the client list.
Brief a colleague properly before you put them on a clock.
WorkflowsSchedule a whole pipeline when the steps must not vary.
Tool groupsWhich capabilities actually do anything on an unattended run.
IntegrationsConnect Slack and email so scheduled output lands where your team works.
Manage schedules programmatically with create_schedule, list_schedules, pause_schedule, and trigger_schedule.