Why an agent registry has to live outside any one company

A bank sets up an AI agent to handle its customers' paperwork. The bank is careful. It lists the agent in its own identity system, names an employee as its sponsor, gives it narrow permissions, and can switch it off from one screen. If you work at the bank, this agent is one of the best-documented things in the building.

One afternoon the agent sends a request to a hospital. A customer wants a surgery slot moved, and the agent has been asked to arrange it.

The hospital's booking desk looks at the request. It can see a name, a message and a patient number. It can't see the bank's directory, because that belongs to the bank. So the desk has a choice between believing the request and refusing it. Neither choice is a good one.

What the big directories do

Several large identity vendors have built ways for a company to register its agents, and what they've published is worth reading.

Microsoft's documentation says Entra Agent ID is now generally available (the page carries a 1 May 2026 date). It lists owners and sponsors for agent identities, so that an organisation can govern them. Okta wrote on 29 April 2026 that Okta for AI Agents is generally available, and that its Universal Directory gives "one place to register AI agents", with a kill switch that applies across connected systems. Ping Identity announced Identity for AI on 6 November 2025, and a 24 March 2026 release says it is generally available.

That is a lot of real work, and we think it's good. A company that runs agents should absolutely have this. Our complaint isn't about any of it.

Where the wall is

Each of those directories answers questions for the company that runs it. The bank's directory knows who sponsors the bank's agent. The hospital's own directory knows nothing about it, and the hospital would be the one acting on the request.

We read the pages above looking for a way for an outside office to check an agent from another company's directory. We didn't find one documented. That's a statement about what we read, not about what the products can do, and we may well have missed something. If you work at one of these companies and we have it wrong, please write to contact@saakshya.org and we'll fix the page and say we did.

The picture is the whole argument. Each company trusts the agents inside its own walls, and nobody can see across. Adding a better directory inside each wall makes the walls tidier. It doesn't put a door in them.

Why the register can't belong to one of them

You could answer that the biggest vendor should just run the shared register. Something like that may happen, and we'd understand why people would want it.

Three things make us uneasy about it, and they aren't all the same kind of worry. A register owned by one vendor is shaped by that vendor's products, so a competitor's agents get second-class treatment even when nobody means it. It can be changed by one company's pricing decision, or switched off by one company's outage. And an office that depends on it has handed the question "who counts as a real agent" to someone with a commercial interest in the answer.

India has run into this shape of problem before. Payments could have been left as a set of bank-owned networks that didn't talk to each other. A shared rail meant a bank, a wallet and a small shop could all check the same thing. We use UPI only as a comparison here. We have nothing to do with it, and we're independent of the National Payments Corporation of India, UIDAI, MeitY and every company named on this page.

What we'd like instead

A register that no single company owns, run by its members under rules published openly. The company directories keep doing what they do. They'd issue credentials that the shared register can vouch for, and check them in return. A hospital could then ask one place about an agent from any bank, and a bank could ask about an agent from any hospital.

We're setting up a non-profit to do that. It isn't incorporated, and we have no members yet. We have a reference implementation and a specification that we're preparing to publish. We don't claim any of the vendors above have agreed to anything, because they haven't.

What we don't know

We don't know whether the directories will want this. A vendor may reasonably prefer to build its own cross-company layer. If so, we'd rather hear it early.

We don't know how many offices outside a company would actually check, instead of accepting the request and hoping. Checking costs time, and a desk with a queue will be tempted to skip it.

We don't know whether a member-run non-profit can stay neutral once real money and real regulators are involved. We have a plan for how to try, and we expect it to change when people who run such bodies read it.

And we haven't read every page these vendors publish. Their products move quickly, and what we say about them is a snapshot from 12 October 2026. For anything that matters to you, read their pages first, and tell us where we're out of date.