Long tasks that keep running
Hand a big job to your agent — an import, a research sweep — and let it keep working in the cloud after you close the app, with a live card to stop or steer it.
Some requests are bigger than one reply: "import last month's runs from these files", "go through these twenty links and summarise each", "check every collection and tell me what's stale". Your agent can hand work like that to a long task — a background job that keeps running on our servers after you leave, and reports back in the same chat when it is done.
How it starts
Just ask. When your agent judges that a request will take more than a few minutes, it starts a long task on its own and tells you so in one line — "Working on Import last month's runs in the background — I'll report back here." You do not have to phrase anything specially; if you want to be explicit, say "do this as a long task" or "take your time on this".
A job card appears under that reply.
The job card
The card shows where the task is right now:
- Status — Queued, Working…, Waiting (between steps), Done, Stopped (you pressed Stop) or Stopped early (it ran out of time or hit a problem).
- Step and time left — how many steps it has taken and roughly how many minutes remain of its budget (20 minutes by default, up to an hour if the agent asked for more).
- Progress lines — the last few things it did, with the tools it used.
- Stop — ends the task at its next safe point. Everything it already did stays.
- Steer — a text box. Type a note ("skip the archived ones", "also check the sleep collection") and send it; the task reads it at its next step and carries on with your note in mind. You can steer as often as you like.
You can close the app, switch phones, or lose the connection — the task does not depend on your screen being open. Open the chat later and the card shows the latest state.
When it finishes
When asking for reusable templates from a source, specify whether you want matching official Library items reused. The agent should preserve source ranges and individual set instructions in the template. Planned targets belong in target fields or notes; actual performance stays blank until you record it. Saving a future plan does not mean you performed it.
A task can also wait until a later observation time. Its card shows Waiting until with the time and reason. While it waits, it makes no AI requests. The wait counts toward its total time budget, so ask for enough time for both observations and the final report—for example, 25 minutes to compare something now and in 20 minutes. You can close the app during the wait. Stop ends a waiting task immediately.
The task writes its final report as a normal message in the same chat, and you get a notification if you are not looking at the chat at the time. Anything it created or changed is listed on the card, and every action appears in the Actions tab of the AI page (in the sidebar, under Trash) like any other agent action.
If the task runs out of its time budget before it is finished, it stops with what it managed, says how far it got, and keeps all of it. Say "continue" in the chat to pick it up from there.
If its report reaches the response length limit, the task shows a failed status and keeps the partial report with an incomplete-answer notice. Saved work stays available. It does not automatically start another paid response; you can ask it to continue.
Ordinary chat requests can also stop when they reach a configured token or estimated-cost limit. The reply says which limit was reached and that work may be incomplete. Completed changes stay saved. These limits are checked between AI steps, so a step already running can pass the limit before its usage is known.
Scheduled long tasks
An automation that runs your agent on a schedule can also run as a long task when its instructions are long, or when you give it a time budget in minutes. Everything above applies — including Stop, which you will find on the automation's row while it runs.
Good to know
- One long task per request. If you ask for two big things at once, the agent will start one and tell you.
- A long task never asks you questions mid-way — it makes reasonable choices and states them in the report. If you want a say, steer it or stop it.
- Tokens are billed as one conversation turn for the whole task; you can see the cost on the AI page.
If a connection retry repeats the same request, it keeps the original background task. If the server stops while handing work off, recovery retries automatically; you do not need to keep the chat open.