Home › AI agent scheduled tasks: what 36 recurring jobs look like across two businesses

AI agent scheduled tasks: what 36 recurring jobs look like across two businesses

Affiliate disclosure: Mulkern AI Systems is an independent affiliate partner of Easy (easy.one), a product of Smart Apps Limited. If you buy Easy through our links, Mike Mulkern earns a commission, and the coupon applied at checkout (MICHAEL) is his. Everything on this page describes how we run our own businesses on Easy. It is our experience, not a statement by Easy.

On 1 November 2026 our Easy installation ran 36 recurring scheduled tasks across 15 agents. Between them they write the morning report, draft social posts, research, audit our sites, and back up a CRM, for both of Mike's companies. This page covers what those AI agent scheduled tasks do, the jobs that watch all the others, and the reply rules that keep their messages worth reading.

What the scheduled jobs cover

Most of the scheduled tasks produce something a person reads. The morning report goes out at 9:00 Central. A research pass runs at 8:00, social drafts are written at 7:00, and an SEO audit covers one property every Wednesday. Others are quieter: a nightly CRM backup, an hourly check on usage against our daily target, and weekly product idea passes on Monday mornings.

Each job runs in its own session with its own brief, and no person watches it run. That is the point of scheduling the work, and it is also the risk. A job can fail, drift, or go quiet, and nobody notices unless something is built to notice.

A health sweep before the morning report

At 8:45 Central, fifteen minutes before the morning report, one job reads the state of the whole installation: the scheduled jobs, token use, services, disk, and every memory index. The memory check has its own step because that failure is silent. Agents keep saving notes that never become searchable, and nothing looks broken on screen.

Two rules came from mistakes. Before acting on a failing job, check when it last ran and whether the cause has since been fixed, because an old error looks the same as a new one. And when several jobs were switched off within the same few minutes, treat it as a person's deliberate choice, not jobs failing on their own. When everything is healthy the sweep writes one line for the report.

An estate watch every four hours

Every failure alert used to go straight to the owner, which made the owner the monitoring system. The rule now is that the agents keep track of every job themselves and bring the owner only decisions. A watch job runs every four hours and reads the job records directly.

One step reads the exact text of each recent message that reached the owner. A delivered message counts as junk if it is narration, a bare status word, or an explanation of which branch an agent chose. Junk teaches a person to ignore the channel, so the watch treats it as a defect in that job's brief.

The watch escalates only decisions, each with a recommendation, what happens if the owner does nothing, and a link to the page where the owner would act. If nothing needs a decision, it sends nothing. Most passes end that way.

A weekly drift check against our own notes

Every Friday at 9:00 Central, one job reads our live software sites and compares them with what our knowledge base says. A disagreement is the finding, whichever side turns out to be wrong. It also tests every checkout link, because a dead link is a dead checkout.

An earlier version re-fetched one page and confirmed its size had not changed. That is a heartbeat, not a check. The current scheduled job reads each page live, follows redirects, and reports. It never edits a live page.

Reply rules that keep the channel worth reading

Our scheduled jobs reach the owner through one shared chat bot, so every message arrives under the same name. Each brief now ends with a delivery tag: the last line of the reply names the job that sent it. The tag goes last because the first words of a reply are read by a machine.

That machine reading matters. Once, a weekly report was recorded as a failure, and raised a false alert, because the reply opened with a line of thinking out loud. The report underneath was fine. So the first word of a reply is the first word of the report.

Jobs built to catch a change reply with a silent token when nothing changed, and that token is not delivered. A healthy week is a short message, and that is a successful result.

What we would tell anyone setting up scheduled tasks

Write the reason for each rule into the brief, with the date it was learned. A scheduled job starts fresh every run, so the brief is the only memory of why a step exists.

Give every job a budget and a stopping point. Keep the watchers read-only, and let a person approve changes: our rule is that changes need a review and Mike's approval before they ship.

In the Easy documentation, see Scheduled tasks and watchers and Stay in control.

Questions

More from how we run on Easy

• An AI chief of staff that writes the daily report for two companies
• Scheduled AI agents with approval gates: the rule we run under
• An AI SEO audit agent that covers one property each week
• See every workflow we run