Skip to content

Agents, workflows and scheduled tasks

What it does

Three building blocks let the assistant work without you typing each request.

An agent is a saved configuration: instructions, a model, tools (web search, URL fetch and connectors) and knowledge areas. You run it on demand from the Try it tab, on a schedule, or from a webhook.

A workflow chains saved agents into a repeatable process, with branches, loops, human approval and delivery to an external channel. It is edited as a draft and then published as a versioned snapshot.

A scheduled task is the simplest of the three: it runs one prompt on a schedule and puts the result into a chat conversation, optionally also by email.

Which one to use:

  • Use a scheduled task when one prompt, repeated, is enough. Example: a daily summary of the news.
  • Use an agent when you want reusable instructions, specific tools or knowledge areas, and optionally a schedule of its own.
  • Use a workflow when several agents must run in order, or when a person must approve a step before the process continues.

Details: Scheduled tasks, Agents, Workflows and Approvals.

How to ask

For a scheduled task, you fill in a form instead of writing a chat message.

  1. Open Scheduled in the left rail and choose New scheduled task.
  2. Enter a Name and the Prompt, the text that is sent to the model on every run. Write it as a complete request that needs no follow-up question.
  3. Pick a Model.
  4. Choose the Schedule (Daily, Weekly or Interval): daily or weekly at a time of day, or at an interval from 1 minute to 366 days. You enter the time in your local time and the interface converts it to UTC.
  5. Optionally enter an address in Also send by email (optional) to receive the result as well.

For an agent, open Agents in the left sidebar and choose Advanced setup in the page header. Give it a name, a purpose, the instructions, and a model. To run it on a schedule, open the Triggers tab of the agent, stay on Schedules, and add a schedule. A schedule can deliver each result to a webhook, a Mattermost channel or an email address. Run now starts a run on the scheduler's next tick, within about a minute. How to write good instructions for an agent that handles incoming mail is described in Process inbound email with an agent.

For a workflow, open the workflow and use its Schedule dialog. Under Trigger, choose Every fixed interval, Once a day (UTC) or Weekly. For Weekly you also pick a Weekday. A time of day in this dialog is entered in UTC, not in your local time, and that includes the weekly time. Choose Save schedule to arm it.

To add a person who must approve a step, add a Human-Approval step to the workflow. You set the approver, what happens if nobody decides in time (end the run or continue automatically), and a timeout from 1 to 8760 hours.

Tips

  • Name each task by its result, for example "Monday sales summary", so the conversation is easy to find later.
  • A run of a scheduled task is appended to the same conversation each time. You see an unread dot when a new result arrives.
  • Start with a short interval only to test. The first run of an interval task starts within about a minute, so you see the result quickly. Then change it to the cadence you really want.
  • Web search is used automatically on a gateway that has it, so a prompt like "summarise today's AI news" can work without any extra setup.
  • Use the agent's Try it tab to check the instructions once before you put them on a schedule.

Limits

  • Scheduled tasks are limited per user. The Custom plan allows 10 active tasks, the free trial allows none, and an enterprise deployment has its own limit. A paused task does not count.
  • Scheduled runs draw on the same prepaid wallet as your chats.
  • Daily and weekly times shift by one hour for you across a daylight-saving change, because the stored time is UTC.
  • A workflow has at most one schedule. It runs as the workflow's owner, and a run is skipped if the previous one is still in flight.
  • A failing schedule's switch stays on, but its timing can change. For a scheduled task or an agent schedule with a fixed interval, every failed run in a row doubles the wait before the next run, up to 64 times the interval and never more than 24 hours; the normal interval returns after the next successful run. A daily or weekly task or agent schedule keeps its time. A workflow schedule with a fixed interval tries at most once every 24 hours after 5 failed scheduled runs in a row (counted from when the latest failed run started); a failed attempt waits another 24 hours, and a successful run resumes the normal interval. To resume at once, fix the cause and save it again (open Schedule, keep Schedule enabled ticked, click Save schedule). A daily or weekly workflow schedule keeps its time. Transient errors of a scheduled task are retried first, up to 3 attempts per run in total. You get at most one failure notice per task in 24 hours.
  • If a run touched protected personal data on a gateway with personal-data protection, the result is emailed only when the recipient is an active registered user of your own organisation (your administrator can switch this off, and then no such email is sent). Otherwise the result is not sent. Mattermost and webhook delivery is always blocked in that case. In every blocked case (email for an agent schedule, scheduled task or workflow; Mattermost and webhook for an agent schedule or workflow) the owner gets a short notice without any content, and the configured recipient gets nothing. A scheduled task's result always stays in its conversation, and an agent schedule's result stays in the Runs tab of the agent.
  • The Triggers (Schedules and Webhooks), Runs and other tabs of an agent belong to its owner and to workspace administrators. A colleague an agent was shared with sees only Overview and Try it.
  • The viewer, demouser and finance roles cannot create these items.
  • A tool hold of an agent that nobody reviews expires and counts as rejected. It can wait 1 hour, 1 day or 7 days (the default). A workflow Human-Approval step instead ends the run or continues automatically when its timeout is reached, depending on the choice you made for that step.

When something goes wrong

  • No new result appears. Nothing needs resuming: a failing schedule stays on. The last run failed, and the cause is in the failure notice you received. Fix the cause. With a fixed interval, the next run can come later than usual after several failures in a row; a workflow schedule with a fixed interval tries at most once every 24 hours after 5 failed scheduled runs in a row. Save the schedule to resume the normal interval at once (see Limits).
  • You cannot create another task. You reached the limit of active tasks. Pause or delete one, or ask your administrator about the limit.
  • The email did not arrive. If the run used protected personal data, the email goes only to an active registered user of your organisation. For any other recipient the result is not sent. Read it in the conversation (scheduled task) or in the agent's Runs tab (agent schedule).
  • The Scheduled entry shows a Premium badge, or the pages redirect to the upgrade page. The feature is not part of the workspace's plan (the API answers 403 feature_disabled), or your role may not create scheduled tasks. Ask your administrator.
  • A run is refused because the wallet is empty or frozen. The run is refused with 429 quota_exceeded, like an interactive chat, with no downgrade to a cheaper model. Top up the wallet, or upgrade, to resume.
  • A run waits and does nothing. It is probably waiting at a Human-Approval step or at a held tool. The approver decides in the Approvals inbox.
  • The run time is an hour off. See the daylight-saving note under Limits.