Unattended automation rewards teams that write the rules down and punishes everyone else. The list below is a professional distillation — ten rules, each drawn from the Ops Atlas entry for AutomationOps, each phrased so it can be adopted, audited, and checked off without a consulting engagement. Nothing here is theoretical; every rule exists because its absence has already produced a documented outage somewhere expensive.
Ten AutomationOps rules for an estate you can defend
1. Put one accountable name on every job
Not a team name, not a rota slot — one accountable human. The Atlas treats ownership as the entry requirement for the whole discipline, and its standing benchmark is 90% ownership coverage across the estate. A job whose owner left the company last quarter is not owned; it is orphaned, and orphaned scripts with admin rights are what governance debt looks like in an incident report. Picking the name is easier than it sounds: whoever notices first when the job breaks is the owner, and the title should appear beside the job in the register, the runbook and the alert routing. Ownership that lives in three places gets corrected in none.
2. Grade every job by consequence before anything else
Four grades suffice. Read-only work sits at grade 0. Reversible change to a single target is grade 1. Reversible change across many targets is grade 2, and it earns rehearsals, batches and health checks. Grade 3 is irreversible or data-destroying, and there the requirements harden: second-person approval, soft deletes, staged rollout. The top grade mirrors NIST’s dual-authorisation pairing — the AC-3(2) / CM-5(4) couplet — so auditors recognise the pattern immediately. Grading is an afternoon’s work with a spreadsheet and honesty. Most teams who try it discover their nightly jobs include at least one grade-3 resident that had been enjoying grade-0 privileges, which is a discovery to make on a Tuesday rather than during an outage.
3. Rehearse with eyes open
Predict before you act. A dry run reports intended actions without performing them, but not all rehearsals are equal: at least one widespread configuration tool leaves command-level probes out of check mode by default, meaning a clean report can conceal the exact step that would have failed. Read what the rehearsal covers, not just what it concludes. Two habits make rehearsals worth trusting: keep the output, because a dated prediction log is the cheapest incident forensics money can buy; and treat any step the rehearsal could not simulate as an automatic candidate for a single-target live trial before the full run.
4. Treat an empty target list as a stop signal
A run handed zero targets must refuse to start — empty never means all, and it never means nobody either. The rule entered industry canon through Google’s Diskerase service, which erased live data after an unrepaired empty entry, and it keeps earning its place through smaller, quieter saves every week. Extend it to surprises of scale: a target list ten times larger than last week’s deserves the same refusal as an empty one, because both mean the job’s assumptions no longer match the world it is about to act on.
5. Cap what any single run may touch
Wide jobs proceed in ordered batches, canary slice first. Tooling exposes this directly — ordered waves with a failure-percentage ceiling on each group — so a bad run damages one small batch while the rest waits, untouched and informed. A ceiling is not pessimism; it is the difference between a correction and a catastrophe. The arithmetic makes the case plainly: a job that touches four hundred targets in slices of four fails at most four targets before a human looks, while the same job in one pass can fail four hundred times before anyone’s morning coffee. Same code, same night, different company-ending potential.
6. Let runs fail small
Agree in advance how many errors a run may collect, then make halting automatic at that count. Two is a workable start. Without a budget, a failing job retries its way through an entire fleet overnight and converts one bug into an archaeology project. The budget also protects the on-call rotation from its worst instinct, which is to keep an eye on a misbehaving run because stopping it feels like giving up. Stopping early is not giving up; it is choosing to debug one failure instead of four hundred.
7. Wire the stop switches, then drill them
Every job needs a documented pause control, and the estate needs one global halt. The Atlas specifies drilling twice a year for grades 2 and 3, and settles the authority question in advance: whoever holds the pager pulls without asking and explains afterwards. A switch that has never been pulled is a rumour, and drills keep paying past the first rehearsal: teams routinely discover that a job feeds a whiteboard, a phone list, or a partner feed, so “stopping” it needs a manual fallback written down in advance. Better to learn that during a quiet Friday drill than at 3 a.m. with an incident channel open.
8. Prefer jobs that seek states, not perform actions
An idempotent job aims at a condition — “ensure running” rather than “start” — so repeating it is harmless. Some actions are inherently once-only — a charge, a shipment, an outbound message — and those instead carry caller-supplied request IDs recorded atomically. This pairing is why Google’s repair jobs could safely repeat every fifteen minutes, and Molecule’s standard test sequence already contains an idempotence step, so regressions announce themselves early. A practical test for any job you inherit: run it twice against a staging target. If the second run makes no change, it seeks a state. If it acts again, it performs an action, and it needs a request ID before it goes anywhere near production.
9. Admit AI agents through the same door as everything else
Autonomous operations assistants have arrived: AWS generalised its DevOps Agent this March (2026), while Microsoft’s SRE Agent already works Azure estates. Their admission rules are not new: the narrowest identity that works, read-only until trust is earned, every action routed through the same guards a script faces. Agents governed separately from automation is how estates end up governing nothing. The enabling pattern already exists: policy as code. When every automation is registered in a declarative file — grade, ceilings, identity, review date — an agent’s proposals route through the same tier-driven approvals and pipeline guard tests as anyone else’s, and the audit trail does not care who, or what, requested the run.
10. Keep a scoreboard, or none of this is observable
Net hours returned per month. Run success rate. Human fallback rate. Induced-incident rate, in the mould of DORA’s change-failure measure. Guard trip rate. Drill-measured time to stop. The count of unregistered automations, which predicts trouble before it happens. Six figures, one page, reviewed quarterly — the estate’s vital signs. The numbers also settle arguments that would otherwise run forever: a job whose net-hours figure stays negative for a quarter is a hobby, not an asset, and the honest response is retirement rather than another tuning sprint. Guard trips belong on the page too, read as receipts rather than failures — every trip is a mistake that got caught by design instead of by luck.
The AutomationOps recap
- Name an accountable owner per job; target 90% coverage.
- Grade by consequence; harden requirements at the top.
- Rehearse changes; read the coverage, not just the verdict.
- Refuse empty target lists; silence is not consent.
- Batch wide runs; canary first; cap the wave.
- Set failure budgets; halt automatically.
- Wire stop switches; drill twice yearly; pager pulls first.
- Seek states, not actions; request IDs for the rest.
- Give agents narrow identities and read-only starts.
- Score the estate quarterly; watch the unregistered count.
Professional operations is not about avoiding automation — it is about being able to answer for it. An estate run by these ten rules can be handed to an auditor, an incoming engineer, or an AI assistant, and every one of them will find the same thing: names, numbers, limits and switches, all current, all defensible. That is the standard this field now sets, and it is entirely reachable from where you are sitting.
