Skip to main content
A scheduled task is a stored prompt plus a trigger. Every time it fires, Cerebrum opens a new chat, works through the prompt on its own, and leaves the result there for you to read.

​
Creating a Task

Ask in chat

“Every Monday at 9, summarize open incidents.” Cerebrum turns that into a schedule and confirms it with you before saving. It can also list, edit and delete your tasks.

Fill in the form

Workdesk → Scheduled tasks → New task. Useful when you want to write the schedule expression yourself, or change one without describing it.
Asking in chat is usually faster, and it is the better option if you would rather not write a cron expression — describe the cadence in your own words and let the agent compose it. You can also paste an expression you already have and ask what it does.

​
Triggers

Schedule

Fires on a cron expression evaluated in the timezone you pick.

Manual

Run now, always available, whether or not a schedule is set.

​
Writing the Schedule

The expression takes 5, 6 or 7 space-separated fields:
Each field accepts an asterisk, a plain number, a list (1,15), a range (1-5) or a step (*/2). Sunday is both 0 and 7. Shorthands @hourly, @daily, @weekly, @monthly and @yearly stand in for the whole expression.
The form shows the next three runs as you type. That preview is the quickest way to check an expression means what you intended.

​
Two Rules Worth Knowing

A year makes the task run once. With no year the expression repeats forever. Add one and the task fires while that year lasts and is then spent — its schedule is dropped, though the task itself stays and can still be run by hand. This is how you say “just once, on this date”.
Day-of-month and day-of-week must both match when you restrict both. 0 9 13 * 5 is Friday the 13th, not every 13th plus every Friday. Some other cron implementations treat this as “either”; Ardor does not, because the scheduler underneath does not.

​
Limits

  • Schedules that fire more often than the platform minimum are rejected.
  • Ten tasks per user.
  • A task whose runs are all in the past is rejected — check the year.
  • After several consecutive failures the task is paused automatically, so a broken task does not keep burning usage. Re-enable it once the cause is fixed.

​
Every Run Is a Fresh Chat

Runs do not share memory, so write the prompt so it stands on its own — the agent brings nothing forward from the last time. Mostly the agent receives exactly your prompt and nothing else. A run is a deferred conversation: you wrote the instructions, so there is nothing to add. One thing is appended, because the prompt could not have carried it — when the last run was. It is absent on a first run, and it is what makes “summarize what changed since the last run” work: the date moves every time, so you cannot write it into the prompt yourself. It comes as a tagged block after your prompt:
The block is addressed to the agent, so the chat hides it — the run’s conversation shows your prompt as you wrote it.

​
Settings

Model

Pick a model for the task, or leave it on the workspace default.

Tool approvals

Checked reviews risky tool calls before they run, exactly as in chat. Full access skips that review — choose it only for a task you are content to leave unsupervised.

​
Watching a Task

The task page lists recent runs with their status, what triggered them, and a link to the chat each one produced. Succeeded and Failed mean what they say; a failed run records the error, so a task that stops working tells you why.