Agent tasks
Recurring runs the agent starts on its own: how to set them up, when they run and how to cap what they spend.
If an agent's instructions are its job description, tasks are what it has standing in its calendar. You do not tell a colleague every Monday that a summary is due either – you agree it once.
Say you want a summary every Monday morning of contracts expiring within a month. You write it to the agent as a task, and from then on it prepares it by itself, without reminding.
Tasks are configured on the agent itself, in the Tasks section, and they are triggered by time.
Putting such a task together
In the Tasks section of the agent settings.
The task name is how you recognise it in the list – for example Weekly contract summary.
In the Instructions field describe what the agent should do on each run. Use @ to add a specific table or record.
Repeat and time. For our example Weekly, Monday, seven in the morning.
The task becomes part of the agent only once you save.
A task brief is not the same as the agent's instructions. Instructions are the job description; a task brief is a specific standing duty in the calendar. The app will not save an enabled task without a brief and writes An enabled task must have concrete instructions.
If the task needs its own material – a template the summary is built into, say – attach it under Files for this task only. The agent gets these files only during that task.
When it will run
You pick the repeat from ready-made options: Hourly, Daily, On weekdays (Mon–Fri), Weekly, Monthly and Monthly on the last day. Hourly adds Interval (hours) between one and twenty-three, weekly adds Days of the week, monthly adds Days of the month – and everywhere else simply a time. The last option suits period closings, because it copes with a shorter February too. Intervals shorter than an hour are not available.
Below the settings runs a Next run line with an estimate of the next start. It really is an estimate; the exact moment is decided by the server. A task also keeps its own time zone, and when it does not match your device's, Datify points that out – so you are not surprised that seven in the morning landed somewhere else than you expected.
If the app were unavailable for a longer stretch, only the last missed run is caught up. So no ten summaries arrive at once after an outage.
Turning off, on and trying out
Every task has a Task enabled switch. A disabled task stays saved with all its settings, it just does not start on its own – you recognise it in the list by the Disabled label. This is handy for seasonal things you do not want to delete.
Rather than waiting a whole week for a result, you can try the task with Run now. You need saved changes though; with unsaved settings the button is inactive with the note Save your changes first. A manual run also does not move the regular slot – Monday's summary arrives anyway.
Enabling and disabling a task is a change to the agent settings, so it creates a new version. Changing the schedule restarts the cycle; renaming a task or editing its brief leaves it alone.
Keeping the cost down
Tasks spend AI credits without your supervision, so they can be capped. In the Spending section turn on Limit automated spending and enter a number of credits per day, week or month. The running state then shows a used of line along with the date the limit resets.
The limit counts automated runs only, meaning tasks. Your conversations with the agent never count towards it, so an exhausted limit will never block your chat. Only an agent manager can set it; an Editor sees it but cannot change it.
Once exhausted, automated runs are paused while the chat keeps working. Skipped runs do not disappear: they stay in the agent's activity with the status Skipped, the reason and the reset date, so it is immediately clear that the task did not fail – it just did not get a budget.
The workspace-wide credit balance is covered by Managing AI credits.
Where to see the results
Finished tasks are in the agent's Activity tab on the Tasks card. Everyone with access to the agent sees them – unlike personal chats, which stay private. Agent activity and versions describes it.