Ask for something online and the server answers with a three-digit number. If all went well, 200. If it can’t find what you asked for, 404. If you’re not allowed in, 403. We almost never see these numbers as we use the web, but with every click, they pass back and forth behind the screen.
Among those numbers, one has sat there for nearly thirty years, holding its place and barely used: 402. Its name is Payment Required. The people who first laid down the rules of the web sensed that a day would come when machines would pay one another, and they left this slot open for it. For years the standard carried only a note beside it: “reserved for future use.” That day was slow to arrive, and 402 became the most famous empty room on the web.
In late 2025, the sleeping number woke. What woke it wasn’t a person but an AI agent roaming the web in a person’s place.
Payment with the person removed
Until now, a person has always stood at the end of a payment. Swipe a card, tap a one-click button, and at the far end of the act there was a human confirmation: yes, buy this. Even the subscription fee that drains from your account each month is no exception. A person agreed once, in advance, that it was fine to keep pulling the money automatically. The machine isn’t deciding, moment to moment, on its own, to pay.
Agentic payments remove the person from that final seat. When an agent, mid-task, decides “to do this I need to buy this data” or “to finish this job I’ll have to rent a server,” it doesn’t ask a person. It pays, on its own, and moves to the next thing.
None of this is entirely new as an idea. Machines paying machines, so-called M2M (machine-to-machine) payments, were talked about back in the Internet-of-Things era. The electric car pays for its own charge; the fridge orders milk when it runs low. But the machine of that era only followed rules written in advance. It was closer to executing an if-statement, “when the milk drops below one carton, order more,” and a person had already decided what to buy and when.
What the agent changed is who does the deciding. Now, what to buy isn’t all settled up front. The agent looks at the situation and decides on the spot. It registers a domain, pays a per-call fee to use someone else’s API, deploys its own output to the cloud, and sometimes hires another agent. That a person can’t know the shopping list in advance is the decisive break from the old M2M.
The three problems to solve
When the person drops out of the end of a payment, all the small things a person quietly did become empty at once. Those gaps fall roughly into three.
The first is trust. From the store’s side, there has to be a way to know whether this agent, now stepping up to pay, really carries its owner’s mandate, and whether it’s buying within the bounds the owner allowed. To a person we posed the question “is this really you?” with a password and a fingerprint. What do we ask a machine?
The second is commerce. There has to be a shared procedure that the store and the agent both understand for the agent to find a product, build a cart, and place an order.
The third is settlement. At some point, money actually has to move from one side to the other. That final transfer has to happen quickly and surely, and without the hassle of opening a new account and registering a card.
Here is what’s happening as of 2026, in one line. No single standard solves all three at once. Different camps have each taken one layer and are pushing their own protocol as the standard. So there isn’t one “agentic payment protocol” but several, and rather than competing head-on, they sit at different layers. Below, we take the three one layer at a time.
Settlement: the sleeping number wakes
At the very bottom, the layer where money actually moves, 402 came back to life. What revived it is a method built by Coinbase called x402.
The mechanism itself is simple. An agent requests some data or function from a server. Instead of handing it over, the server answers with 402 and includes the payment details: “to use this, pay this much, here.” The agent pays on the spot, then attaches proof that it settled and sends the same request one more time. This time the server checks the proof and returns 200 along with what was asked for. Request, 402, payment, retry. The whole transaction closes in a single round trip.
What’s worth noticing here is the procedure that disappeared. No need to create an account, no need to be issued an API key, no need to register a card. The gate of sign-up and registration that a person routinely passes through when first using a service vanishes entirely. To an agent, each of those gates was an obstacle that summoned a human’s involvement, and x402 erases it by folding the payment itself into the HTTP request. That’s why it fits best with the picture of “an agent paying a few cents on the spot every time it uses an API.”
There’s a foundation beneath this ability to settle on the spot without signing up. x402 doesn’t rest on bank accounts or the card network but on digital money on a blockchain — that is, on stablecoin payments. Instead of splitting authorization and settlement across days the way a card does, the value moves on the spot and it’s done. Coinbase built x402 on top of exactly this immediacy.
In practice, x402 is also the one running the most. Its second version arrived in December 2025; in February 2026, Stripe integrated it into its own payment network, and Cloudflare supports x402 transactions as well. That said, since the method is still young, it hasn’t accumulated much scrutiny yet. Academics have already published a paper analyzing how an attacker might bypass x402’s payment and pull resources without paying. Having folded payment into an HTTP round trip, the research probing the seams of that round trip has begun to attach itself in turn.
Commerce: when the chatbot closes the sale
One layer up, the commerce layer where products are found and orders placed, was claimed jointly by Stripe and OpenAI. It’s the ACP (Agentic Commerce Protocol) the two companies put out in September 2025.
What this layer defines is the procedure by which an agent and a store trade. It ties together, into one standard flow, everything from the agent discovering a product to initiating payment to completing the purchase. What stands out is that this didn’t stay at the concept stage. The checkout flow where you buy something mid-conversation inside ChatGPT is the first case of ACP running in a live service. From the store’s side, it’s closer to opening one more standard counter, next to the one that serves human shoppers, to serve agent shoppers. Google then jumped into this layer with a separate standard of its own, UCP (Universal Commerce Protocol), reaching more broadly from product discovery all the way to returns.
Trust: turning the mandate into a document
At the very top, and perhaps the trickiest layer, is trust. Google claimed this one with AP2 (Agent Payments Protocol).
The core idea of AP2 is the mandate, a kind of authorization put in writing. The authority a person grants an agent is recorded not in words but in a digitally signed document. There are two kinds. One captures intent: “I entrust that, on these conditions, you may buy,” something like “buy a ticket on this route within a budget of five hundred thousand won.” The other is a signature on the actual, finalized cart: confirmation that this is the item the agent chose, at this price.
Why these documents matter shows up when something goes wrong. When an agent buys the wrong thing, or buys at a price it never should have, there has to be a basis for asking, “was this within the mandate?” AP2 leaves a chain of signatures running from the person’s authorization to the agent’s execution, so that the chain can be traced back afterward. It doesn’t move money directly. It handles only “who authorized what.” Of the three problems above, filling the first gap, the gap in trust, is this layer’s job.
The card networks step in too
It isn’t only the newly arrived camps that have entered this arena. The card networks, which have carried people’s payments for decades, have put out answers of their own. This is the part most interesting to those who know the day-to-day of payments.
Visa announced its Trusted Agent Protocol in October 2025. It pairs a verified agent identity that Visa issues with a separate consent record signed by the bank that issued the consumer’s card, so that a store can confirm the agent is genuine and mandated. The notable design choice is that this rides not on some entirely new technology but on the signed HTTP messages the web already uses. That lets a store establish trust without laying down a whole new system.
Mastercard takes a slightly different tack. Under the name Agent Pay, it extended the card-tokenization technology the company has long used. It binds three things into a single token: which agent, up to which store scope, and under which consent policy. As a result, an agent like ChatGPT or Copilot can complete a payment without ever once holding the real card number. Spending outside the bound scope is refused by the token itself. To the earlier question, “who authorizes the agent’s spending,” this is an answer that bakes the policy into the token.
Payment splinters, authority converges
Laid out like this, it looks like a chaos of proliferating standards. Yet by mid-2026, the current has been settling in one direction. Who does the settlement, and by which procedure the commerce runs, splits into several strands, but the grammar of authority, “who authorized what,” alone converges on one place. That place is Google’s AP2.
For now, though, it’s still at the stage where each party has declared a direction and floated a plan. Mastercard has said it will expose AP2’s mandate concept directly in the process of issuing its own tokens, and Visa has said it will merge its two-pronged protocol into a single developer toolkit and make it compatible with AP2 by default. Standards fights don’t always play out as announced. Even so, a current is taking hold in which the execution of payment scatters down many roads, x402 or the card networks, while the language for confirming whether that payment stands on a legitimate mandate converges on AP2 alone.
This is no accident. Settlement and commerce can have several methods coexist and the market still runs. It’s the same as being able to buy things with a card or with an on-chain payment. But trust alone, without a common language, leaves no one comfortable enough to transact. So whoever holds this layer’s standard comes to hold the chokepoint of the agent economy. On the surface it’s a quiet fight over a mandate spec, but it’s really a struggle over the most valuable seat in the house.
When the buyer isn’t a person
Once these three layers settle into place, an old question returns wearing a new face. Who sets the price? Until now, price tags were written for human eyes. Whether it feels expensive, whether you’re in the mood to open your wallet right now, was the ground the price stood on. When the buyer becomes an agent, that ground shifts. An agent doesn’t hesitate over a mood; it judges only by conditions and budget, and it spends without rest. Just as the flat-rate subscription, which propped up heavy users on the strength of an average, wobbled first in front of it, the entire way of setting a price gets rewritten against a judgment that isn’t human.
And we should also see what the things being built right now are leaving empty. What the protocols answer is mostly “how do you pay safely,” not “who keeps watching that spending.” You can set the bounds with a mandate and cap the limit with a token, but tracing an agent’s spending, hundreds of items a day, the way a person combs a statement, is another problem entirely. A person could point at one line and say “I didn’t spend this,” but an agent’s spending is too fine and too plentiful to point at. Authorization and oversight are different things, and what’s being hurriedly built right now is mostly the authorization side.
In the place where the person was pulled from the end of the payment, one seat remains that no one has yet filled. It’s the seat of someone who undoes a bad purchase, notices money leaking, and, at some moment, says “this isn’t right.” The protocols are quickly deciding how an agent pays, but who guards whether that money was spent properly is still left blank.