AI agent registry

An AI agent registry is the inventory an organization keeps of every AI agent it builds or runs. Each entry records who owns the agent, which business process it serves, which model and version it uses, which tools and data it may reach, who approves its consequential actions, and whether it is in testing, live, suspended or retired.

The registry answers the first questions in any review or incident: which agents exist, who is accountable for each, and what each one may do. Access reviews, risk assessments and incident response all start from it. An agent with no current entry cannot have its permissions recertified, and an operator cannot be sure that a kill switch reached every instance.

The internal registry and the EU database are separate records. Under Article 49 of the EU AI Act, providers register high-risk AI systems listed in Annex III in the EU database set up by Article 71, and deployers that are public authorities, or act on their behalf, register their use of those systems. A bank's internal registry usually covers many agents below the high-risk threshold, and it holds operating details such as tool permissions and approval owners.

Keep the registry tied to the systems that run the agents. An entry kept by hand in a spreadsheet drifts as versions change, while an entry updated by the release process stays current. In one of our own test environments, the registry grew to 17 agents, about 14 of them near-duplicate test versions that nothing retired, which made the live agent hard to find. Record every version change with its date and approver, so that an auditor can match any past action to the agent version that took it. Write the agent identity and version into each decision record when the record is created, because a sealed record cannot be amended later.

In KLA, the Agent Registry records each agent's owner, versions and release state, and the Tool Catalog records the tools each agent may call.