Sylvain Kalache’s piece names something most AIOps dashboards will happily hide: your average MTTR can fall while your worst incidents get slower. AI clears the routine cases — inspect alerts, form hypotheses, correlate deploys, ship the fix — and the routine cases are exactly how engineers built the intuition they need for the incident automation can’t touch.
- 🎯 The metric lies by aggregation — median MTTR drops, tail MTTR on novel high-sev incidents climbs, and a single average buries the trade
- ⚠️ Never-skilling is the sharper risk than deskilling — junior responders may never build the pattern library senior on-call quietly runs on
- 🔍 Bainbridge’s 1983 “ironies of automation” called this: automate the easy 95% and humans inherit only the hard 5%, with less practice for it
- ⚡ Aviation’s fix is deliberate manual practice — pilots still hand-fly for exactly the conditions autopilot can’t handle
- 💡 Treat responder practice as an SLO, not a side effect — if the agent handles everything, schedule the drills the real incidents used to be
The production question isn’t “should we let AI run incidents” — for routine ops that’s already the default. It’s whether your observability and on-call design deliberately keep humans engaged enough to stay sharp for the tail. The HN discussion is full of engineers who’ve already felt that gap widen. If your agent resolves 95% of incidents, who on your team is still practicing for the 5% that pages the CEO?