Skip to content Skip to footer

An agent’s identity is not enough to get the truck moving

In brief. Recognizing who sends an order is only the first check. The delegation has to say how much the agent can spend and what to do if the confirmation is missing. In November 2026 the Linux Foundation welcomed Agent Name Service, seven specifications for recognizing an agent.

A supplier receives a reorder from an agent of a regular customer. The terms are the agreed ones, the delivery address is new. The agent identifies itself correctly. The company still has to decide whether to load the truck.

The scene is hypothetical, but it contains a distinction that will become very concrete: knowing who is talking to us does not clarify everything they are allowed to do. An employee can genuinely belong to the purchasing office and have no authority to change a destination. The same goes for a program acting on the company’s behalf.

On November 10, 2026 the Linux Foundation welcomed Agent Name Service, an open standard for identifying and finding agents through the domain name infrastructure and certificates. The project is backed by eleven organizations; the specifications are organized in seven documents. They address identity, registration and verification, and do not replace the company’s decision on how much to delegate. Linux Foundation · Announcement

The value of a common infrastructure is avoiding that every relationship has to start from scratch. The supplier recognizes a verifiable counterpart, can tell one agent from another and build a service around that information. But the next step belongs to buyer and seller: what commitment can that counterpart take on?

When the task widens

In a talk on leadership I described a test run by a friend who works in consulting. He had entrusted an agent with finding small businesses whose website could be improved, preparing a new version and contacting the potential customer. When a hotel asked for new photographs, the work also involved finding a photographer through my friend.

It is his account, not a process I verified step by step. The detail that interests me is the request for photographs. The initial task concerned a website; to complete it, the system ran into a further need. At that point autonomy stops being an abstract function and becomes a question about commitments, people and money.

An agent can have permission to ask a professional about availability. Confirming an assignment is another matter. It can present a price without being able to negotiate it, collect a quote without being able to accept it. These are differences an organization has to make understandable to the counterpart as well, before anyone starts working.

Back to the truck: a useful delegation can allow reorders within a spending cap, with approved suppliers and for sites already registered. The change of address falls outside that perimeter. It could be legitimate, but it needs further authorization. The supplier must be able to verify the mandate, and whoever grants it must be able to revoke it.

Stopping without always asking

If every variation goes through a person, though, we have built an assistant that generates approval requests, sometimes faster than we can read them. The saving risks disappearing in the time needed to read them. It is a real problem, especially as the number of operations grows.

The answer is choosing where to leave autonomy. A change in quantity within agreed limits can be ordinary. A new destination or a different payee can require a check. The quality of a delegation also shows in how many correct operations it lets through undisturbed, while keeping visible the ones that really change the risk.

Limits also have to be read as a whole. Many small orders, each below the authorized threshold, can exceed the budget for the period. The mandate therefore has to consider commitments already made, not assess each request as if it were the first.

This work starts from the exceptions in the existing process. Whoever authorizes a variation today knows what information they need. Recovering it and making it usable by the system is often more useful than a generic rule such as “ask if in doubt.” The machine’s doubt and the manager’s doubt do not necessarily coincide.

The order without confirmation

Then there is a situation in which identity and authorization are both correct, but the decisive piece of data is missing. The agent sends the order. The connection drops before it receives the confirmation. The supplier may already have accepted it. Pressing “Retry” can produce a second load.

What is needed here is the ability to recognize the same operation through a stable reference and look up its outcome. The procedure has to distinguish a failed attempt from a result not yet known. If the check does not clarify what happened, the case goes to a person with the information to reconstruct it. Simply asking “do you want to retry?” transfers the uncertainty without resolving it.

The same criterion applies when the mandate is revoked: you need to know which operations have already gone out and which can still be stopped. Switching off an agent does not automatically cancel what it asked a supplier to do.

For whoever runs the company, the autonomy contract therefore also covers recovery from errors and interruptions. Before increasing the number of delegated orders, it is worth watching one incomplete order, following its reconstruction and understanding who has the last word.

The truck can leave when sender, mandate and the outcome of previous operations are clear. A verified identity is a good start. The goods move on the basis of everything else.

Read next The test comes before the code · AOV: three letters and two ways of running a company

Image: Linux Foundation