An agent registry is the system of record for an organisation’s AI agents. Each entry names the agent’s business owner, the task it performs, the tools and data it may touch, the model and prompt versions it runs on, its risk tier and its latest evaluation results. Changes to any of those fields go through the registry, so the entry always describes what is running in production.
It is change management for software that acts. A human change board meets weekly; an agent’s behaviour can shift overnight when a model, a prompt or a permission is changed. Without a registry, a firm cannot answer the questions a regulator or auditor asks first after an incident: which agents could have done this, who approved their access, and what changed last. DBS Bank’s CEO describes its agent controls in exactly those terms: the registry, accountability, observability, traceability and evaluations.
The rule of thumb is simple: if an agent is not in the registry, it does not get credentials. Build the registry before the tenth agent, not after the hundredth.