Global AI Assurance · Layer 02
Agent Identity Registry
An agent that can act inside a regulated institution is an institutional entity. It needs an identity, an owner, a purpose, a jurisdiction and an authority envelope, on the same terms as a person holding a mandate. Anything less and the institution cannot say who did what.
The registry record
Every field is here because a supervisor, an auditor, a resilience test or an incident review asks for it.
Agent ID
A durable institutional identifier, independent of the platform hosting it.
Named human owner
One accountable person. Not a team, not a mailbox, not a vendor.
Purpose
The single business purpose the agent was approved for, in operator language.
Jurisdiction
The supervisory regimes that bind the activity the agent performs.
Model and version
Behaviour changes with the version. Assurance has to be version-aware.
Tools
Every tool the agent may invoke, declared before it runs.
Permissions
Scoped authority per tool rather than access or no access.
Data access
Classes of data reachable, with the transfer basis for each.
Financial authority
The value ceiling above which the agent must stop and escalate.
Action authority
Which actions are recommend-only and which may be executed.
Escalation authority
Who receives the escalation, and within what tolerance.
Evidence route
Where the agent's actions and refusals are written, and for how long.
A sample register
Bounded sample. The unassigned entry is deliberate: it is what discovery usually finds first, and it is why the registry exists.
Treasury liquidity agent
Position reads and funding recommendations
- Owner
- Group Treasurer
- Jurisdiction
- UK, EU
- Autonomy
- Recommend only
Sanctions screening assistant
Alert triage and narrative drafting
- Owner
- MLRO
- Jurisdiction
- UK, US, SG
- Autonomy
- Human-gated execution
Third-party review agent
Vendor questionnaire analysis
- Owner
- Head of TPRM
- Jurisdiction
- EU, CA
- Autonomy
- Bounded execution
Client correspondence drafter
Regulated communications drafting
- Owner
- Head of Compliance
- Jurisdiction
- US
- Autonomy
- Human-gated execution
Settlement exception agent
Break investigation and routing
- Owner
- Head of Operations
- Jurisdiction
- EU, JP
- Autonomy
- Bounded execution
Model documentation agent
Validation dossier assembly
- Owner
- Head of Model Risk
- Jurisdiction
- US, UK
- Autonomy
- Recommend only
Vendor-embedded summariser
Unapproved feature in licensed software
- Owner
- Unassigned
- Jurisdiction
- Undetermined
- Autonomy
- Bounded execution
Why identity comes before control
Attribution
Without an agent identity, an action is attributed to a service account and the accountability chain breaks at the first question.
Revocation
Authority you cannot name is authority you cannot withdraw. Identity is what makes an intervention possible.
Evidence
Every refusal and every permitted action is written against the identity, which is what turns oversight into a record.
Identity establishes who the agent is. The permission fabric establishes what it may do.
Agent Permission Fabric