Patrick Parker has written the most rigorous account of agent governance the field has produced this year. Read it with the label removed and it is something he does not quite claim it to be: the case for zero standing privilege at the point of action.
A paper is circulating under the title The Laws of AIdentity, written by Patrick Parker of EmpowerID. It is an homage to Kim Cameron's Laws of Identity and an extension of Martin Kuppinger's coinage, AIdentity. It is also — and this is meant as the highest compliment the imprint can pay — the clearest single piece of architecture written about AI agents to date. The failure modes are named and distinct. The lineage to capability security is acknowledged rather than quietly absorbed. The standards inventory is accurate and load-bearing. This is the work of an architect, not a marketing department.
Which is exactly why it deserves to be read precisely. Because once you do, a question becomes hard to set down: why is it called identity at all?
There is a simple test for whether a word is doing architectural work in a document or merely flying the flag over it. Remove the word. Replace it with a plainer one. See whether anything has to change.
Take Parker's seven laws, strike the prefix, and read what each one actually instructs a system to do:
Says: intent, authority, credential custody, execution and accountability are separate actors, and a system that collapses them loses attribution. A statement about evidence and attribution — not about the agent possessing an identity.
Says: authority is granted before the specific action exists, so you must authorise the action, not the actor. Govern the verb, not the noun. The CASS house position, stated in Parker's own words.
Says: authority must be contextual, time-bounded, purpose-scoped and revocable — "the agent may persist; its authority should not." That is a definition of zero standing privilege. It does not mention identity at all.
Says: authorise across the action lifecycle, not at the front door — a control loop, not a checkpoint. A statement about when control happens, located at the action.
Says: the safest credential is the one the agent never holds; request the outcome, not the secret. The elimination of standing credential custody. A privilege primitive, not an identity one.
Says: every tool, model and sub-agent in the path must prove provenance and integrity, or you inherit the risk of the least governed participant. Supply-chain integrity at the point of action.
Says: prove the action was authorised, constrained and attributable — receipts, not logs. Prove the action, not the login.
Seven laws. Authorise the action. Bound and expire the authority. Never let the agent hold the credential. Verify the chain at the point of use. Prove what happened. Now rename the paper The Laws of Delegated Action. Nothing changes. Not one law, not one failure mode, not one architectural principle has to be rewritten. The only casualty of the rename is the word that gave the paper its title.
The identity word is the one part of the paper that does no architectural work.
This is not a reading imposed from outside. Parker's own structure concedes it. In his mapping back to Cameron, he is candid that the inheritance is asymmetric: four of the seven laws are new for AIdentity — Split Actor, Generated Intent, Continuous Authorization, Proof-Carrying Action. Those four are the genuinely novel content of the paper. And not one of them is about identity. They are about action, time, custody and evidence.
The laws that are inherited from the identity tradition fare differently. Directed Identity and Consistent Experience — Cameron's identity-native dynamics — are demoted to "cross-cutting" considerations rather than laws in their own right, because at the action layer they do not carry weight. The paper's own scaffolding tells the story plainly: where it innovates, it has left identity behind; where it leans on identity, it has nothing new to add.
That is the signature of a frame that has outlived its usefulness. The engineering forced Parker to govern the action — because you cannot govern runtime-generated intent any other way — and the vocabulary he inherited made him file the result under the actor.
It would be easy to call this a quibble. It is not. CASS has argued from its first paper that identity is a directory — a supplier of authentication and authorisation inputs — and nothing more; that treating it as the control plane is the category error from which standing privilege, and therefore breach, follows. The AIdentity framing reintroduces exactly that error at the moment the field can least afford it.
Call the discipline identity and you invite the whole apparatus back in: the agent must be enrolled, issued, made a first-class identity, governed as an actor. But every control in Parker's paper that does real work governs the action and isolates the credential — that is, it works precisely by refusing to trust the agent's standing as an actor. The receipt in Law 7 does not need the agent to be an identity. It needs the action to be attributable. Those are different requirements, and conflating them is how a sound architecture acquires an unsound centre of gravity.
The paper's expansion to drones, vehicles and medical devices is its most ambitious move and its most revealing one. A safety envelope for an autonomous vehicle — geography, speed, proximity to humans, emergency-stop rules — is not an identity claim by any available definition. It is a bound on action. Bundling it under "identity" is the same instinct that elevated identity to a cosmological principle in the first place: the urge to make identity the container for everything. The container is action. Identity is one input to it.
None of this diminishes the paper. The opposite. Parker has independently derived the architecture CASS calls governing the action rather than the actor, arriving at it from the identity tradition rather than against it, and stress-testing it across a wider surface than this imprint has attempted. Where our forthcoming work on agent access control places the privilege primitive at the point of execution — the endpoint as the control plane the directory never was — Parker's seven dynamics populate that structure with a rigour worth studying. On the substance, there is very little daylight between us.
The daylight is in the name. And the name matters, because names decide which industry shows up to build the thing. Call it identity and the identity industry builds it — with directories at the centre and the agent enrolled as an actor. Call it what the seven laws actually describe — delegated action, governed at the point of action, with no standing authority and no held credential — and a different, better system gets built.
Patrick has written the strongest argument yet for governing the action rather than the actor. He has filed it under the actor. Reclaiming it is a single edit: strike the word, keep the laws.
Strike the word. Keep the laws.
New Floor Reports, papers, and the occasional correction to the consensus — sent direct, single voice, no vendor cc.