Skip to content

Meet your AI team. Introducing Experts.

Meet the Experts
← All articles

How to Run AI Agents on Autopilot: 7 Rules for Routines

The short answer: an AI agent runs well on autopilot when its job is narrow and its report is predictable. Give it one deliverable per run. Fix the order of the report. Let it stay quiet when nothing changed, and never let it repeat itself. Ban invented facts. Cap what it can spend. Then, after it goes live, keep checking that it ran, what it did and what it cost.

The examples below are quoted from AutoGPT's public skills catalog, the files that define each AI Expert in our marketplace, as of November 18, 2026.

What "on autopilot" should mean

Autopilot is for work that comes back: the Monday pipeline review, the morning meeting brief, the Friday competitor scan. A routine is that job plus a schedule. The agent returns to it while you are away, and the result waits for you to review.

It does not mean nobody is flying. You choose the job, you read what comes back, and you decide what happens next. If you are still choosing which jobs to hand over, start with how to automate repetitive tasks with AI agents.

Eight of our routines

ExpertRoutineRuns
Max, Account ExecutivePerfect-fit leads every MondayMondays
Frankie, Executive AssistantToday's meetings, and what still needs youEvery weekday
Maria, SEO Content ManagerWhat ships this week, and what is stuckEvery weekday
Daniel, Financial AnalystMonday budget pace checkMondays
Nadia, Market Research AnalystWhat competitors shipped, said, or changedFridays
Remy, Email Marketing ManagerHow last week's lifecycle email actually didMondays
Robin, Customer Support SpecialistEscalation and SLA watchEvery weekday
Sofia, RecruiterUrgent thread checkHourly on weekdays

They were written for different jobs, but the same seven rules keep showing up.

Seven rules for a routine that runs itself

1. One deliverable per run

The best titles say what arrives and when: "Perfect-fit leads every Monday." You can tell from the title whether it ran.

Most routines are slower than you would guess. Of the 46 in our skills catalog, 30 run weekly, 11 run every weekday and 3 run monthly. One runs twice a day and one runs hourly. Pick the slowest schedule that still gets you the answer in time.

2. Fix the order of the report

Maria's routine spells out how to report:

"report in this order: what ships in the next seven days with the owner on each line, what is late and by how far, what is stuck waiting on one person or one missing proof point, and what has no owner or no ship date."

The same order every run means you can read it in seconds. Frankie's morning brief goes one step further and asks for "a one-line summary at the top for anyone with thirty seconds."

3. Let it stay quiet

A report that arrives every day with nothing in it trains you to stop reading. So the routines are told when to say almost nothing:

  • Maria: "If nothing ships this week, nothing is late, and nothing has changed since your last run, say so in one line and stop."
  • Daniel: "When nothing is off pace, send one line with the count of lines checked and stop."
  • Sofia, who checks every hour: "If nothing new endangers a loop, say nothing."

4. Never the same flag twice

The second way to lose a reader is to repeat yourself. These routines check their earlier reports before they write a new one:

  • Maria: "Never flag the same stuck row two runs running unless it got worse."
  • Nadia: "Never brief the same change twice."
  • Sofia: "a thread gets at most one flag per day."

5. No invented facts

A routine that runs unattended must be able to say "I don't know."

  • Maria: "Never invent an approval, a draft, or a date."
  • Nadia: "No change without a link, and never pad it to look busy."
  • Daniel: "If the data has a gap in its dates, name the missing dates, do not extrapolate over them, and ask for that export in one line."

6. Start with setup questions and a spending cap

A routine cannot guess where your data lives. Before its first run, each one asks a few setup questions. 38 of the 46 ask exactly two, nearly always where to read and when to report. Maria's are "Where is your editorial calendar?" and "What time should this land, and in which timezone?"

Then set a limit. In AutoGPT, each marketplace Expert's work is capped at $5 a week until you change it.

7. Watch it after it goes live

Software teams do not launch something and walk away. They watch it run, because that is how they catch it going off the rails. A routine needs the same step: a file moves or a login expires, and the report still arrives on time.

For the first few runs, and every few weeks after that, ask three questions:

  • Did it run? A quiet routine and a broken one look the same from your inbox. Daniel's routine, with nothing to report, still sends "the count of lines checked".
  • What did it do? Read the record, not only the report. Alex, the Product Manager, is told: "Log what you read and what you could not reach." Nadia's routine logs every page it fetches, "including the ones that failed."
  • What did it cost? A jump in cost means the routine is doing more than it used to.

In AutoGPT, you can open an Expert's activity and follow the work. When a run breaks, the Expert's status says why: which app disconnected, or the known cause of a run that could not finish.

When an answer surprises you, change the routine before its next run.

A routine you can copy

Fill in the brackets and give this to your agent:

Every [Monday at 8:00], read [where the data lives] and your last report. Report in this order: [first], [second], [third]. One line per item, each with its source. If nothing has changed since your last report, say so in one line, name what you checked, and stop. Never repeat a flag unless it got worse. Draft [the follow-up] and send nothing. Never invent a number, a name or a date. Close with anything you could not reach or verify.

Run it by hand twice before you schedule it. The first run shows you what the agent is missing. The second shows you whether it remembers the first.

What to keep off autopilot

  • Anything that leaves your company. Emails, posts and payments go out after you approve them. In AutoGPT, every marketplace Expert asks for your approval before acting externally.
  • Decisions. A routine hands you the number. You decide what to do about it.
  • Jobs whose answer rarely changes. Run those monthly, or ask when you need them.

How this works in AutoGPT

Hire a ready-made Expert or create your own, give them the context and connect your tools. When a job needs doing again, set up a routine: agree on the details and the schedule, and the results wait for you to review. You can change or pause the schedule at any time, and set a spending control for each Expert.

Browse the Experts to see who does what.

Frequently asked questions

What does it mean to run an AI agent on autopilot?

It means the agent does a recurring job on a schedule without being asked each time, such as a lead list every Monday or a meeting brief every morning. You still set the job up, review what comes back, and approve anything that leaves your company.

How often should an AI agent run on a schedule?

As often as the answer changes, and no more. Of the 46 routines in AutoGPT's skills catalog, 30 run weekly, 11 run every weekday, 3 run monthly, one runs twice a day and one runs hourly during working hours. Start weekly and speed up only if you keep wanting the report sooner.

Is it safe to leave AI agents running unattended?

It is when the routine only reads, drafts and reports. Keep sending, paying and editing behind your approval, cap what each agent can spend, and read the first few runs before you trust the schedule.

What does it cost to run AI agents on a schedule in AutoGPT?

Scheduled work draws on your balance. Each marketplace Expert's work is capped at $5 a week until you change the limit, so a routine cannot run up a bill while you are away.

See what an Expert can take off your plate

Tell Otto what needs doing, or choose an AI Expert and give them the context for the job.

Start free for 7 days