Book a demo
Topics · GOVERNANCE-AGENTS

Asset Mgmt Agent Identity Setup | Moring

Setting up AI agent identity at an asset management firm means establishing the identity architecture before agents scale: per-agent credentials issued from a system of record, machine identities kept separate from human ones, key-bound authentication, scoped access packages, and a mediation point where every call can be checked and logged.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Setting up AI agent identity at an asset management firm means establishing the identity architecture before agents scale: per-agent credentials issued from a system of record, machine identities kept separate from human ones, key-bound authentication, scoped access packages, and a mediation point where every call can be checked and logged.

How it works

Setup decisions are the ones that are expensive to reverse, so the ordering matters more than the tooling. Curity's guidance supplies the first and most consequential decision: agents should stay machine identities, never blended with human identities, because mixing the two creates ambiguity that is hard to maintain. The consequence for an asset management firm is structural: an agent is not a user with a service account, and the identity system must represent agents as a distinct population with its own lifecycle.

The second setup decision is credential type. Curity prefers asymmetric, key-bound credentials, signed tokens and certificates with mutual TLS, over symmetric secrets and static API keys, and the reason is verifiability: a key-bound credential can prove its holder is the party it was issued to, which matters when credentials are ephemeral and agents act without sessions. Shield's architecture describes the issuing function formally, answering who the agent is, acting for which human, running which build, and then sender-constrains the holder on each use.

The third decision is where checks live. A mediation point between agents and the systems they reach, such as an MCP gateway, authenticates the agent as itself, checks every tool call against policy, logs immutably, and can pause within seconds, per Lutril's design. IBM's agent gateway description frames the same layer as a centralized control point that routes requests, manages permissions, and logs activity. Setting up per-integration enforcement instead means every new system is a new gap; setting up the mediation layer once means enforcement scales with the population.

The fourth setup element is the access package structure. Orchid's guidance describes ownership, least privilege, access packages, lifecycle automation, and the intent-to-execution loop as the working pattern, and setup is when those structures are created: an entitlement bundle per agent tied to its registered purpose, reviewable as a unit, and reassessed when the role changes. Setup is also when the baseline Microsoft's Cloud Adoption Framework recommends is defined, a centralized, enforceable governance and security floor every agent satisfies before operating.

The fifth element is discovery and registration of what already exists. Nudge Security's guidance reports most enterprise environments already have agents operating without IT approval, deployed by employees, auto-provisioned by platforms, and extended through MCP server connections, so setup includes a sweep of OAuth grants, API keys, MCP connections, and non-human identities, converting found agents into registered ones. An asset management firm that skips this step builds its architecture for the agents it plans and leaves the ones it has uncontrolled.

The regulatory frame reaches setup because the controls are hard to retrofit credibly. SR 11-7 expects model risk governance commensurate with impact, and the identity architecture is part of the control set for agents on investment data. SR 26-2 third-party guidance applies where identity infrastructure or agent platforms run with vendors, so setup includes deciding what the firm can independently verify and direct. The EU AI Act, for firms operating in the European Union, expects documented oversight, and the setup record is the first document.

A practical setup sequence, synthesized: establish the agent identity system of record; issue per-agent key-bound credentials with named owners; define access packages against registered purposes; deploy the mediation layer with logging and pause capability; run the discovery sweep and register what it finds; define the baseline and lifecycle automation; and retain evidence from the first action onward. Each step is cheaper now than it will be after the population grows, which is the whole argument for doing setup deliberately.

Frequently asked questions

What are the first architecture decisions when setting up agent identity?

The first is that agents remain machine identities, separate from human identities, because Curity's guidance finds that blending the two creates ambiguity that is hard to maintain. The second is credential type: asymmetric, key-bound credentials such as signed tokens and certificates with mutual TLS, preferred over symmetric secrets and static API keys for verifiability. Shield's architecture adds the issuing function that answers who the agent is, acting for which human, running which build, and sender-constrains the holder on each use.

Why place enforcement at a mediation layer rather than per system?

Because one mediation point covers the whole agent population, while per-integration enforcement means every new system is a new gap. Lutril's MCP governance design shows the layer authenticating the agent as itself, checking every tool call against policy, logging immutably, and pausing within seconds, and IBM's agent gateway definition describes the same centralized point routing requests, managing permissions, and logging activity. Setup is the moment to make this choice, since retrofitting enforcement across growing integrations is where gaps accumulate.

What is an access package and why does setup need one?

An access package is a bundle of entitlements tied to one agent's registered purpose, reviewed as a unit rather than as accumulated individual grants, and reassessed whenever the agent's role changes. Orchid's guidance lists it among the working patterns of agent identity governance alongside ownership, least privilege, lifecycle automation, and the intent-to-execution loop. Setting this structure up early means Microsoft's Cloud Adoption Framework baseline, a centralized enforceable floor every agent satisfies before operating, can reference well-formed bundles instead of ad hoc permissions.

How does setup handle agents that already exist in the environment?

Through discovery and registration: a sweep covering OAuth grants, API keys, MCP server connections, and non-human identities, per Nudge Security's guidance, which notes most enterprise environments already contain agents operating without IT approval through employee deployments and auto-provisioning. Found agents are converted into registered identities with owners and scoped packages. For an asset manager, SR 26-2 third-party guidance and the EU AI Act both expect the firm to know and document what it operates, so the sweep is itself part of the compliance record.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu