The Registry and the Record

Governments are beginning to require that an agent be registered before it runs. A national regime became enforceable this month. A bill in the US Senate would put certain user-facing agents on a federal register before they may reach the platforms they act on.

Registration establishes that the agent exists and what it was declared to do. It is silent on what any single action the agent took was allowed to do.

The register is a provisioning-time object.

A registry entry is written once, before the agent acts. It fixes a declaration. It names the responsible party. It names the agent identity. It names the class of task the agent is filed to perform, and in the stronger regimes a risk tier. Every field in it is a statement of intent, recorded in advance.

The run is where intent meets contingency. The agent composes inputs the filing never saw. It calls tools on results that did not exist when it was registered. It takes an action partway through the run that the declared purpose neither authorized nor forbade in any specific term, because the declared purpose was a category and the action is a particular.

A registry says the agent is allowed to operate. It does not say a single action the agent took was governed.

Three moves, one list.

Three developments this year make the same move. China's implementation opinions on intelligent agents became enforceable on July 15, 2026, with filing obligations in high-risk sectors and a tiered decision-authorization framework. A bill before the US Senate would require certain user-facing agents to register with the Federal Trade Commission before they may access the large platforms they operate against. An international standards body has opened a focus group on trust and identity for agents across the agent lifecycle.

Each of them puts the agent on a list before it runs. The move is sound as far as it reaches. A regulator that cannot enumerate the agents operating in its jurisdiction cannot supervise them. Enumeration is a precondition for supervision. It is not the supervision.

Registration answers which agents exist and who is answerable for them. Supervision answers what a given agent did, under what authority, at 14:32 on a Tuesday. The register cannot answer the second question. The second question is the examination.

A declared purpose is a category, not an action.

A filing binds an identity to a declared purpose and a responsible party. That is necessary and it is not enough. A declared purpose is a category. An action is a particular that falls inside it or does not, and the filing was written before any particular existed.

Take an agent filed to assist with credit servicing. It is registered correctly whether a given action waived a late fee, escalated an account to collections, or moved money between two accounts. The filing does not separate those actions because it predates all three. The category covers each of them and governs none of them.

The examiner does not doubt the agent was registered. The examiner pulls one of those three actions and asks what authority bound it, what the agent was permitted to do at that step, and whether the action was reduced or stopped before it ran. The filing holds a declared purpose. It does not hold that answer.

A roster is not an account.

Records-retention and traceability obligations are about conduct. They ask a firm to show what a system did and to reconstruct the decision behind an outcome. A registry proves the agent was declared. It does not prove what the declared agent did. When a framework asks for the evidence behind a decision, a registry entry is upstream of the act and answers a different question.

SR 26-2 took effect on April 17, 2026 and placed agentic AI outside its scope, pointing institutions back at their own risk practices while an interagency RFI is still forthcoming. EU high-risk record-keeping and traceability obligations sit at December 2, 2027 for standalone systems and August 2, 2028 for embedded ones. A national filing regime tells a regulator that an agent is in operation. Those obligations ask what the operation did, decision by decision.

Registration is the roster. The record is the account. A roster of the agents in the room is not an account of what any of them did.

Where the register and the record meet.

The two objects have to be bound. Every runtime decision record has to carry the registration identity, so that a decision pulled in year three resolves to the agent the regulator filed. The register points at the agent. The record proves the action. Bound together, a single pulled decision resolves to a registered identity and to a signed account of what that identity did.

The binding is not automatic. A registry maintained by a regulator and a decision record produced by a vendor live in different systems and different hands. The agent identity written into the record has to be the same identity on the register, and it has to be verifiable by a party who trusts neither the vendor nor the operator. A record that names an agent the register cannot confirm, or a register entry with no record beneath it, leaves the examiner exactly where a filing regime was meant to move past.

What we are building.

Wayfinder Systems Group builds the runtime layer the register points at. Every decision and every learning event is sealed onto a tamper-evident chain at the moment it happens, and each record carries the agent identity a regulator can resolve against its filing. The register establishes that the agent exists. The chain establishes what the agent did, under what authority, and whether the action was reduced or stopped before it ran. The reviewer reads exceptions. The examiner reads a record that resolves to a registered identity and cannot have moved since it was written. Patents held in The Wayfinder Trust. We call her Velma.