How often does an unattended job actually fail?
Nobody publishes this, so here are our own numbers: 4,104 runs by 70 scheduled agents over 30 days, with every alert broken out by kind. Measured, not estimated, and rebuilt from the production database every week.
The rate is boring. The kinds are not.
0.6 percent of runs ended in failure: 25 out of 4,104. If you were expecting worse, that matches most fleets, and it is also why people underestimate the problem. A crash is loud, rare and easy to catch. It is not what costs you a night.
Those 4,104 runs produced 60 alerts in total. Here is what they were:
| Kind | Count | What it means |
|---|---|---|
| FAILED | 25 | exited non-zero, or the client reported failure |
| DRIFT | 20 | cost or duration moved away from this job's own baseline, while nothing failed |
| MISSED | 12 | the run never started: dead scheduler, expired auth, crash before the first line |
| NO_EVIDENCE | 2 | reported success, but the file, URL or assertion it promised was missing |
| STALLED | 1 | started and then stopped reporting before it ended |
Read that table from the bottom up
The rarest one is the one nothing else catches. 2 runs reported success while the file, URL or assertion they promised was not there. Every heartbeat monitor in this category would have counted those as green, because the job did check in. It just did not do anything. That is the failure mode people only discover downstream, when someone asks where the data went.
The second most common one has no category. 20 alerts were drift: cost or duration moving away from the job's own baseline while nothing failed at all. No exception, no missed run, no red status. A job that quietly starts taking four times as long is the earliest warning you get, and almost nothing watches for it.
Only 12 were the one everybody monitors. A run that never started is what a dead man's switch is for, and it is the minority of what actually goes wrong.
What these numbers are not
This is one fleet: 70 scheduled jobs that run a data business, mostly small scrapers, refreshes, health checks and report builders, plus a handful of LLM jobs. The median run takes about a second. A fleet of long agent sessions would fail more often and differently, and anyone who tells you there is one industry failure rate is selling something.
Two more honest limits. The record is built from what each client reported, so a client that lies about its own cost produces a faithful record of the lie. And a job nobody registered produces no alerts at all, which is the one failure mode monitoring can never see.
Why we can publish this at all
Because the runs behind it cannot be edited afterwards. Each finished run is hashed, each day of hashes is sealed into a Merkle root, and each root is chained to the day before and anchored in Bitcoin. So this page is not a marketing claim about our own reliability, it is a number you can recompute: verify a real run yourself, no account needed.
The dashboard version of the same measurement, with the median time to alert and the proof chain, is on in numbers. The fleet itself is live on a real fleet.
If you want your own numbers
Wrap one job and you have a baseline within a week. Free for 20 agents, all detectors on, no card: get a free key.
Figures from the production database on 2026-09-12, over the preceding 30 days. TEST alerts excluded. Nothing on this page is typed in by hand.