Asset Mgmt Agent Identity How Does It Work | Moring
AI agent identity works by issuing each agent at an asset management firm a verifiable credential tied to its registered purpose, checking every action against scoped permissions at runtime, and recording the result. The pattern repeats across issuance, delegation, per-call authorization, and revocation, which is what keeps autonomous access explainable.
Short answer
AI agent identity works by issuing each agent at an asset management firm a verifiable credential tied to its registered purpose, checking every action against scoped permissions at runtime, and recording the result. The pattern repeats across issuance, delegation, per-call authorization, and revocation, which is what keeps autonomous access explainable.
How it works
Mechanically, an agent identity system does four things, and it helps to name them. Shield's agentic identity architecture describes them as issue, sender-constrain, authorize, and record: who is this agent, acting for which human, running which build; is the holder of the credential the party it was issued to; is this specific action permitted; and what did the agent actually do. An asset manager does not need that vendor's implementation, but the four functions are the anatomy of any workable design.
Issuance comes first. Every agent receives its own credential rather than borrowing a developer's account or a shared service key, and the credential encodes the agent's registered purpose, owner, and permitted scope. Curity's guidance argues agents should remain machine identities, never blended with human identities, because mixing the two creates ambiguity that is hard to maintain, and recommends asymmetric, key-bound credentials such as signed tokens or certificates over symmetric secrets and static API keys. OpenA2A's AIM architecture similarly registers each agent before it submits a single action.
Authorization is where the design earns its keep. AWS's Well-Architected Agentic AI Lens distinguishes two patterns an asset manager will meet immediately: an agent acting explicitly on behalf of a user, bounded by that user's permissions, and an agent acting autonomously on a schedule, an event, or another agent's trigger. The second pattern is the dangerous one, and the lens prescribes least privilege, separation of agent and human permissions, and strong authentication for agent identities. Authorization checks happen per call, not per login.
Per-call checking is necessary because agents change behavior mid-flight. Orchid's guidance explains that unlike static service accounts executing predefined operations, agents interpret intent and take variable actions at runtime, so the intent-to-execution loop must reconcile each action with the agent's stated purpose before it lands. Omada adds the failure this prevents: agents can chain decisions, delegate across systems, and stay inside existing permissions while still acting beyond what anyone intended.
Recording is the fourth function and the one auditors see. Every issuance, scope change, check, and denial is logged with the identity that performed it, forming the decision evidence Omada's framework calls for. Dataiku's survey of more than 800 data leaders found only five percent report AI output is traceable all the time, which is the measurable consequence of skipping this function. For an asset manager, the record is what connects a portfolio screening or client report back to the identity and authority behind it.
The regulatory frame gives the mechanics their weight. SR 11-7 expects model risk governance commensurate with impact, and the identity mechanism is part of the control set governing agents that touch investment data. SR 26-2 applies where the identity platform runs with a vendor. The EU AI Act, for firms operating in the European Union, imposes documentation duties that per-call logs and issuance records satisfy directly.
How the loop behaves in practice at an asset management firm: a screening agent issues a credential at registration, receives a scope covering the specific research database and workflow, presents the credential on every call, is checked against policy including who owns it and what purpose it serves, and its denials as well as approvals are retained. When its role ends, revocation is immediate because the credential is independent of any human account. That is the whole machine, repeated across every agent the firm runs.
Frequently asked questions
How does an agent identity get issued and verified in practice?
At registration the agent receives its own credential encoding its purpose, owner, and permitted scope, rather than borrowing a developer account or shared service key. Curity's guidance recommends keeping agents as machine identities separate from human ones and using asymmetric, key-bound credentials such as signed tokens or certificates over static API keys. Shield's architecture describes the follow-on functions as sender-constraining the holder, authorizing each action, and recording results, while OpenA2A's AIM design registers and scores agents before their submitted actions are honored.
What happens at authorization time when an agent calls a system?
Authorization is evaluated per call rather than per login, because agents interpret intent and take variable actions at runtime, unlike static service accounts with predefined operations as Orchid's guidance explains. AWS's Agentic AI Lens distinguishes the on-behalf-of-user pattern, bounded by the user's own permissions, from autonomous execution, which requires least privilege, separation from human permissions, and strong agent authentication. Each call is checked against the agent's scope and stated purpose, and both approvals and denials are recorded as decision evidence.
Why is per-call checking necessary rather than periodic review?
Periodic review evaluates an identity's standing; per-call checking evaluates its behavior. Omada's guidance documents the difference in outcome: agents can chain decisions, delegate across systems, and remain inside existing permissions while acting beyond what anyone intended, so only a check at action time catches the drift. The recorded results also supply the traceability Dataiku's survey found missing at most organizations, with only five percent of more than 800 data leaders reporting full traceability of AI output.
How does this mechanism support an asset manager's regulatory posture?
The mechanism produces exactly the artifacts regulators and examiners ask for: issuance records tying each agent to an owner and purpose, per-call authorization results, and an immutable action log. SR 11-7 expects model risk governance commensurate with impact, and these controls govern the agents that touch investment data, while SR 26-2 applies where the identity platform runs with a vendor. The EU AI Act imposes documentation and transparency duties on firms operating in the European Union that per-call logs and issuance records satisfy directly.