The AI Ops Checklist
Ten checks to run before AI touches production — the ones that separate a demo from something you can sleep on. No fluff, no vendor pitch.
Get the checklist (free)
Drop your email and it unlocks instantly on this page. We'll also send the occasional practical ops note — unsubscribe anytime.
No spam. One click to unsubscribe.
Before you ship
- Name the consumer of every queue. A queue with nothing draining it looks identical to a broken product. Write down who picks up each job.
- Bound every retry. A retry limit that can't actually reach its number isn't a limit — it's a rumor. Test that it stops.
- Make failures verifiable. For non-idempotent writes, on error re-read the intended state ("does it exist?") instead of guessing whether it happened.
- Classify errors by whether a retry can help. "Network blip" and "port already in use" must not share a bucket.
- Keep secrets out of the transcript. Pipe any command that touches credentials through a redactor — and test the redactor against a real field name.
- Back up off the box, and prove the restore. A backup on the same host is a copy, not a backup. Restore one file on a schedule.
- Alert on readiness, not elapsed time. A health check that runs before boot manufactures false alarms and trains you to ignore it.
- Cap the blast radius of an agent. Every action reviewable; the human keeps the final send. Scope by action, not by system.
- Watch for drift, not just crashes. Model output degrades quietly. Review a sample weekly against the tone you intended.
- Write the runbook at 2am, then test it sober. If a second person can't execute it, it's a habit, not a process.
Want the boring parts owned for you? We run exactly these systems at hub.sqs.chat.