Skip to main content
Tasks let you run work without chatting: schedule an agent to run a fixed instruction on a cadence — or once, at a set time — and review each run’s result afterwards. Open the page from “Tasks” in the sidebar. A task is made of three things:

​
Creating a task

1

Open the form

Press “New task” on the Tasks page. Suggestion chips offer common starting points.
2

Describe the work

Enter a title, write the instruction the AI should follow on each run, and pick the agent to run. Leave “Reference past history” unticked unless each run should pick up where the previous one left off (see “Runs and review” below).
3

Set the schedule

Choose “Recurring” or “One-time.” For a recurring task, describe the cadence in words — “every weekday at 10:30,” “every 15 minutes” — and the AI converts it into a cron expression, or fill in the five cron fields (minute / hour / day of month / month / day of week) directly. The form shows the resulting cadence in plain language along with the next run time, so you can check it before saving; if your phrasing can’t be read as a schedule, the form says so and leaves the fields for you to set. Pick the timezone the schedule follows. A one-time task takes a single “Run at” date and time instead.
4

Create

Press “Create.” The task appears in the list with its schedule and next run time.
A task’s instruction runs unattended — the AI cannot ask you mid-run what you meant. Write it so it stands alone, the way you would brief a colleague who works while you are away. You can always ask the AI in a chat to help you draft it.

​
Runs and review

Each firing produces a run whose reply is captured as a normal chat thread — open a run from the task’s history to read the result, and continue the conversation there if you want follow-ups. By default every run starts in a new chat, so one run never sees what an earlier run said or produced. Tick “Reference past history” on the task (in the create form or in Edit) and each run instead continues the task’s previous chat, with that conversation as context — useful for work that builds on the last result, at the cost of a conversation that grows with every run and can make runs more expensive. Only the task’s creator can change this setting.
  • The history in the task’s detail is paged 20 runs at a time, newest first, with “Previous” / “Next” controls
  • Failed runs are flagged on the task’s row so they are easy to spot. A run that could not even start is recorded in the history too, with the reason. “Not run: …” rows name a cause that clears on its own — the agent is archived (the task fires normally again once it is unarchived), or the run failed to launch; “Paused: …” rows name a cause that also disables the task — the agent was deleted, it is no longer shared with this workspace, or the task’s creator is gone or no longer a member
  • “Run now” fires the task immediately, without waiting for the schedule
  • If a run cannot start because the organization has reached its monthly usage limit, or because the task has reached its rolling 24-hour usage limit — counted across all of the task’s chats, and for a task that references past history also per reused chat — the task retries automatically when usage becomes available again. The same happens when the previous run has not been confirmed finished (it is still running or waiting for an answer): the next run is skipped and the task resumes once it ends. The task’s row shows “Last run skipped” with the reason until then, and “Run now” says the same if you press it meanwhile

​
The Tasks page

The tasks are listed on the left — title, agent, schedule, last run, and next run — and selecting one shows its run history and instruction in the column beside the list. When the window is too narrow for both, the detail opens as a drawer over the list instead. Drag the divider between the two to resize the list; the choice is remembered on that device.

​
Editing, pausing, and deleting

“Edit” opens in place inside the detail, with a back control to return.
  • The title, schedule, and timezone can be changed by the task’s creator or a workspace owner
  • The instruction, the agent, and “Reference past history” can be changed only by the task’s creator — a run executes as the creator, using their connector access, so nobody else can change what runs on their behalf
  • Changes apply from the next run onward. Runs already in the history were produced by the previous instruction
  • A task can be toggled between enabled and disabled without deleting it, and deleted outright (deleting stops its scheduled runs and cannot be undone)
Two things can happen without you doing them. A task disables itself when it can no longer run as configured — its creator left the workspace or their account was removed, the agent is no longer shared with them, or the agent was deleted (the task is paused at its next run, which is what records the “Paused: the agent was deleted” row). Fix the cause — or, for a deleted agent, point the task at another one — then re-enable it. If the creator’s organization membership cannot be confirmed, runs are skipped (“Last run skipped” on the row) and the task is paused with a “Paused” badge once that has lasted 24 hours. And for a task that references past history, changing its agent starts a fresh review thread: the next run opens a new chat instead of continuing the old one, though earlier runs stay listed and readable.
If there is no “Tasks” entry in your sidebar, the feature is not enabled in your environment. The AI can create a task from a chat (below), but it cannot change, run, pause, or delete one — those happen on the Tasks page.

​
Creating a task from chat

You can also ask in any chat — “run this every Monday morning,” “remind me next Friday” — and the AI proposes the task on a “Task proposed” card: the agent, the title, the schedule in plain words with its timezone (or the date and time of a one-time run), the next run, and the full instruction. Press “Create” to save it; it starts at its next run time, and the card’s “Open the task” link takes you to it on the Tasks page. Nothing runs at the moment you approve, and the task can be edited or deleted here like any other (In-Chat Forms and Suggestions).
A task created from chat runs as you, with your connector access, exactly like one created on this page. Read the instruction on the card before approving — it is what runs unattended.