Six months into a typical AI rollout, we ask the same question: can you tell us what your agents do and who owns them?
The answer is almost always no.
There is a list somewhere of the tools that were procured. There may be a Slack channel where people share prompts. There is certainly a budget line. But a registry — a structured, maintained record of what agents are running, what they have access to, what decisions they are involved in, and who is accountable for their behavior — almost never exists.
This is not a niche compliance concern. It is a foundational operating risk.
What an agent registry is
An agent registry is exactly what it sounds like: a catalog of the AI agents a firm operates, with enough structured information to understand each one’s scope, access, ownership, and governance status.
At minimum, each entry should capture:
What it does. A clear description of the workflow the agent participates in, written for someone who did not deploy it. Not “AI assistant for the M&A team” but “reviews incoming NDA drafts for non-standard clauses and flags them for associate review before any partner sees the document.”
What it has access to. Data sources, systems, file permissions, external APIs. What the agent can read and what it can write.
Who owns it. A named human being who is accountable for the agent’s behavior. Not a team, not a function — a person.
What its evaluation status is. Has anyone reviewed whether it is performing correctly? When? How often? Against what criteria?
What its escalation path is. When the agent encounters something it cannot handle, what happens? Who gets notified?
Why this is harder than it looks
The challenge is not building the registry. The challenge is that agents proliferate faster than governance. A team stands up an agent to help with a specific task. It works. Word spreads. Someone else adapts it. A third person connects it to a new data source. Six months later, there are twelve variations of the original agent running in different corners of the firm, some of them with access the original deployer never intended.
Without a registry, none of this is visible. The agents are running, but no one has a complete picture of what they are doing or who is responsible for any of it.
The registry is not just a documentation artifact. It is a forcing function. Building it requires someone to ask, for every agent: who owns this? what does it actually do? what can it access? If the answers to those questions are “unclear,” that is a governance failure — and it is far better to discover it in a registry exercise than in a client incident.
When to start
The right time to start an agent registry is before you need one. The second-best time is now.
If your firm has been running AI tools for more than three months, you almost certainly have agents that no one has a complete picture of. The registry exercise is not a heavy lift — it typically takes two to three days to conduct the initial inventory for a team of fifty. It is the maintenance discipline that matters: updating the registry when new agents are deployed, when access changes, when ownership transfers.
Firms that build this habit early find it valuable beyond governance. It becomes a resource for understanding what AI the firm actually has, what it is doing, and where the opportunities for deeper deployment exist. The registry is not overhead. It is knowledge.