Visa's April 2026 agentic transaction rules define an "Agentic Transaction" as an e-commerce transaction undertaken by an Agentic Payment Provider on behalf of a cardholder, based on "cardholder-defined payment instruction(s)," without direct cardholder-merchant interaction. Before the agent can act, the provider must obtain cardholder consent to provision a token, state the instruction's expiration date, obtain acknowledgement of responsibility, and verify the cardholder's identity. The provider must retain this agreement for the duration and produce it on written request.
The developer documentation makes the mechanism concrete. A purchase intent carries an instructionId and mandates[] fields including a mandateId, description, effectiveUntilTime, and a decline threshold amount. The word mandate is doing real structural work here. It is a data object specifying what the agent may do, for how long, up to what limit, and under whose authority.
Scope, duration, ceiling, principal. A notarized letter of delegation would have carried these same elements in 1950. Visa rebuilt them as API fields.
Separately, Cloudflare's July 2026 policy classifies AI bot behavior into three categories: Search, Agent, and Training. Starting September 15, 2026, new domains will default to blocking Training and Agent bots on pages displaying ads, while leaving Search allowed. The classification is purpose-based. Not what the bot does technically, but why it is there and what it intends to do with what it finds.
Purpose classification is ancient institutional logic. A broker could act on a client's behalf within the scope of their license, but not beyond it. A visitor to a corporate office was assessed at the reception desk: Are you here to sell something? To deliver something? To inspect something? The classification was carried by context, by credentials, by the social legibility of the person standing at the counter. Cloudflare is encoding the same principle into HTTP-layer defaults, applied to automated actors that have no face and no counter to stand at.
These two organizations arrived at their designs independently, solving different problems in different industries, and the convergence is the interesting part. Visa needed to know whether a transaction was authorized and by whom. Cloudflare needed to know whether a visitor's purpose was compatible with the site owner's interests. Neither was consciously reconstructing paper-era governance. Both ended up doing exactly that.
The convergence likely reveals something worth naming. Institutional authority does not survive format changes intact. It has to be re-derived. And each re-derivation makes visible what the previous format was quietly carrying, because each generation of institutional furniture has less ambient context to rely on. Visa's mandate fields spell out what a power of attorney implied. Cloudflare's purpose categories formalize what a receptionist assessed at a glance. The paper form didn't need to specify these things explicitly because the humans operating within it already understood the institutional context. The form could be simple because the people were complex.
The newest institutional furniture is the most carefully specified precisely because the actors it governs carry the least institutional context of any actors these systems have ever encountered. That is probably the right direction. But notice how much institutional knowledge has to be made legible, has to be written down as data fields and classification schemas and retention requirements, before a system can govern actors that cannot read the room. Someone always pays the cost of lost context. In this case, the cost shows up as engineering effort, as regulatory complexity, as the painstaking work of converting what everyone used to simply know into something a machine can be held to.