What should a scheduled AI job do when it runs late?
Scheduled AI jobs run late more often than anyone plans for. A late run is fine. A late run that pretends it was on time is not. The job should write down when it was due and when it actually ran, then cover everything since its last good run, not just 'the last 24 hours'.
One of my agents writes an end-of-day summary every night. What happened, what moved, what's waiting on me.
This week it was scheduled for 23:03 on 12 November. It actually fired at 10:03 the next morning. Eleven hours late.
Nothing got lost. The note opened with one line that said exactly what had happened: "Late run: scheduled for 12 Nov 23:03 IST, fired 13 Nov 10:03 IST. Covers everything since the 11 Nov sync (23:10 IST)."
That is the whole essay. A scheduled AI job that runs late should do two things: write down when it was due and when it actually ran, and cover everything since its last good run rather than a fixed window like "the last 24 hours." Late runs are normal. A late run that quietly pretends it ran on time is how work falls through the gap.
Why do scheduled AI jobs run late?
Because the schedule is a request, not a promise.
The machine was asleep. The scheduler was restarted. A queue backed up behind a slower job. A rate limit kicked in. The model provider was slow. None of these are exotic. If you run agents on a schedule for long enough, every one of them will happen to you.
Most setups are written as if they won't. The prompt says "summarise today." The query says "last 24 hours." The note is titled with today's date. All three quietly assume the job ran when it was supposed to.
What goes wrong when a job runs late and doesn't say so?
Two things, and both are silent.
The window slips. Do the arithmetic on my late run. If the job had looked back a fixed 24 hours from 10:03 on 13 November, it would have started at 10:03 on 12 November. The last good run ended at 23:10 on 11 November. That leaves almost 11 hours that no summary would ever have covered. No error, no alert. Just a hole.
The reader is misled. A summary dated the 12th, written with the morning's numbers, looks like the night before. Anyone acting on it, a person or another agent, is acting on the wrong picture of time. That matters more for agents than for humans, because agents read each other's notes literally.
A late run is fine. A late run that pretends it was on time is how work disappears without anyone deleting it.
How should a job know what work it missed?
Remember the last time it succeeded, and start from there.
This is old advice from the infrastructure world. The systemd timer docs describe an option called Persistent=: when it's on, "the time when the service unit was last triggered is stored on disk," and the service is triggered immediately if it would have run "at least once during the time when the timer was inactive." The docs say plainly that this "is useful to catch up on missed runs."
Apache Airflow builds its whole model around the same split between "when was this for" and "when did it run." Its Dag Runs documentation explains that each run has a data interval, and that for a data-interval schedule the logical date is the start of that window, "not when the Dag is actually executed." It also warns that if a pipeline is "not limited to the interval, but instead to Now," you should turn catch-up off, because rerunning it would cover the wrong data.
That warning is the agent version of my bug-that-didn't-happen. "Since now minus 24 hours" is a window tied to Now. "Since the last good run" is a window tied to the work.
- WAKEJob firesOn time or not. Assume not.
- STAMPDue vs ranWrite both times at the top of the output, in the reader's time zone
- LOOK BACKSince last good runRead the last success timestamp, not the clock
- COVERThe whole gapSummarise or process every item in that window, once
- SAVENew last-good markOnly after the output is written, so a crash doesn't skip a window
Should a late job catch up, or skip?
It depends on whether running twice can hurt.
Google's SRE book has a chapter on exactly this trade-off, Distributed Periodic Scheduling with Cron. The authors write: "In general, we favor skipping launches rather than risking double launches," because "recovering from a skipped launch is more tenable than recovering from a double launch." They also point out that many cron jobs, "like a payroll run or sending an email newsletter, are not idempotent," and that owners "can (and should!) monitor their cron jobs."
For agents, I split the work the same way:
- Reading and summarising: catch up. An end-of-day note, a digest, a report. Running late and covering a longer window is safe. Missing a window is the failure.
- Sending and spending: skip, then tell a human. Messages, posts, payments. If the slot passed, don't fire a stale action and don't fire it twice. Record that it was skipped and why. I wrote about why an agent's outward actions need their own rules in The agent needs a yes.
The mistake is using one policy for both. Catching up on a summary is good hygiene. Catching up on a batch of outreach is how someone gets the same message twice.
What should the late-run note actually say?
Three things, at the top, before anything else.
- 01DueWhen the job was supposed to run, in the reader's time zone.scheduled time
- 02RanWhen it actually ran. If the gap is big, say 'late run' in plain words.actual time · delay
- 03CoversThe exact window it processed, starting from the last good run.from · to
It costs one line. It means the next reader, human or agent, can trust the numbers without investigating. It's the same idea as The agent needs a heartbeat, one level down: a heartbeat tells you the job ran. The due-versus-ran line tells you whether it ran when you think it did.
What breaks if you only plan for on-time runs?
Anything that depends on a date getting noticed.
Across my daily notes from 2 to 11 November, the same sync kept flagging 69 rows in one of my trackers that had no follow-up date at all. Its note said it plainly: nothing will surface them. That isn't a late-run bug, but it's the same family. Work that relies on a date or a window to come back up will quietly vanish when the date is missing or the window slips.
The fix in both cases is not a smarter model. It's making the gap visible: a timestamp that says when, a window that says what was covered, a flag that says what nothing will catch.
So what's the rule?
Assume the job will run late. Stamp due versus ran. Cover everything since the last good run. Catch up on reading, skip and report on sending.
Open your scheduled agents today. If any prompt says "today" or "last 24 hours," change it to "since the last good run."
- What should an AI agent do when it gets stuck?Nov 11, 2026 · 7 min
When an AI agent hits a blocker it can't fix, like an expired login or a failed tool, it should stop, record the reason where humans look, ask for one specific action, and resume only on confirmation. Guidance from OpenAI, Anthropic and Google SRE, plus a real paused job.
- The agent needs a retroOct 30, 2026 · 7 min
AI agents don't learn from their mistakes on their own. Write each miss as one rule, decide which rules go live, and put them where the agent reads first.
- The agent needs a heartbeatOct 18, 2026 · 8 min
Production agents rarely fail loudly. They go quiet, or report work they never did. Give each one a heartbeat, a run receipt and an alarm.
Owned, undistributed, and published only here — no newsletter, no funnel, no upsell.