Visa and Mastercard Shipped Agent Payments on the Same Day. Only One Said Guaranteed.

Both networks announced agent payments on 10 June. Visa built a better decision. Mastercard used the word guaranteed — and that is the whole difference.

Share
Visa and Mastercard Shipped Agent Payments on the Same Day. Only One Said Guaranteed.

Fourteen months ago the two networks opened their rails to agents two days apart. This time they did it on the same day, and the gap between them finally has a word in it.

On 10 June 2026, Visa announced its agentic and stablecoin programme at the Visa Payments Forum. On the same day, Mastercard launched Agent Pay for Machines.

I wrote in May of last year that the two networks had opened their rails to agents within forty-eight hours of each other and that neither had said who pays when an agent gets it wrong. They have now done the simultaneous announcement again, with a year of work behind it, and the symmetry has finally broken.

One of them used the word guaranteed.

What each one shipped

Visa. Visa Intelligent Commerce is the umbrella: trust, controls and connectivity so agents can discover, initiate and complete transactions. Underneath it sit several pieces worth naming individually.

Agent Score, built with New Generation, helps a merchant assess whether an AI agent can navigate its site, understand it, and complete a task. Agentic Directory lists the agents and merchants Visa has verified as legitimate participants. A strategic collaboration with OpenAI puts Visa payments inside agentic commerce across OpenAI's surfaces. A Large Transaction Model, trained on billions of transactions, is aimed at fraud detection, authorization performance and false declines. And from Visa's Crypto Labs, a command line proof of concept that lets agents pay for digital services in the terminal using Visa tokenized credentials — explicitly an early concept, not a product.

On the back end: Tokenized Deposits, a technology layer letting banks turn deposits into programmable money while the funds stay on their own balance sheets. Expanded stablecoin settlement pilots across regions, chains and currencies. And stablecoin-linked cards, with more than 160 programmes live or in development.

Two numbers from that list are worth more than the rest of the announcement combined, and I will come back to them.

Mastercard. Agent Pay for Machines extends the Agent Pay programme introduced last year, and it targets a specific transaction shape: high-volume, low-value, low-latency payments between agents and other automated systems. Some of them, the release says, are only fractions of a cent.

It has four parts. Credentialing: each agent gets a credential, and through something called Verifiable Intent, agents can be recognised and transact across ecosystems. Permissioning: organisations set authorization rules and spending limits that are enforced programmatically. Transacting: verified participants connect and transact across providers and systems. And Settling: settlement is guaranteed across multiple rails.

More than thirty organisations are named as initial participants. Among them: Adyen, Coinbase, Stripe, Cloudflare, Global Payments, OKX, Solana Foundation, Tempo and Utila.

Decide versus pay

Put the two lists side by side and they are not the same kind of product at all.

Almost everything Visa announced on the agentic side is about knowing. Can this agent use the site. Is this agent verified. What does the token's history say. What does a model trained on billions of transactions think of this request. Agent Score, the Agentic Directory, enriched token data, the token assurance signal, the Large Transaction Model — every one of those is an input to a decision about whether to approve.

That is an authorization posture, and it is the posture a network takes when it intends to improve the quality of the decision and leave the consequences of a wrong decision where they currently fall.

Mastercard's four verbs end somewhere else. Credential, permission, transact, settle — and settlement is guaranteed across multiple rails.

That is a different sentence. Guarantee is the most expensive word in this industry. It is what separates a network from a messaging standard, and it is the thing the agentic protocols published by technology companies have conspicuously never offered. Nobody building ACP, AP2 or x402 has said they will make anyone whole.

I want to be careful not to over-read a press release. Mastercard does not define the guarantee, does not say against what, does not state who bears a loss when an agent transacts outside its mandate, and does not publish the rule. A guarantee without a rulebook entry is a marketing word. But it is the first time either network has put that word next to agent-initiated money movement, and it is the direction I have been arguing the industry has to go since April of last year. Credit where it is due.

A network routing around itself

Now the detail in Mastercard's announcement that nobody will write about.

Some of these payments are fractions of a cent. A card transaction cannot be a fraction of a cent. The cost floor of a card authorization — the message, the fraud evaluation, the clearing record, the dispute right that attaches to it — is a fixed cost, and fixed costs are fatal below a certain ticket size. This is not a Mastercard problem; it is arithmetic, and it is why micropayments have failed on card rails for thirty years.

So how does a card network offer sub-cent payments?

It does not. It routes them somewhere else. Mastercard says settlement spans cards, accounts and stablecoins, and the participant list tells you which part is doing the sub-cent work: Coinbase, Solana Foundation, Tempo, OKX, Utila.

What Mastercard has built, then, is a credential and permission layer sitting above rails it does not own, for the transaction class its own rail cannot price. That is a serious strategic move and I think it is the right one, but it should be named accurately. The network is competing by not being the rail.

It is also exactly the arrangement that makes the guarantee interesting and hard. Guaranteeing settlement on your own rail is a solved problem with sixty years of practice behind it. Guaranteeing settlement across somebody else's chain, where finality is probabilistic and the asset is issued by a third party, is a new question. Mastercard has either solved it, priced it, or papered it. We should ask which.

The two numbers

Back to Visa, and the figures that matter more than any product in the announcement.

Visa says it has moved billions of dollars in stablecoins across VisaNet, at an annualized run rate of approximately seven billion dollars as of March 2026. And it says issuing banks already settle onchain seven days a week, while Visa is working to extend seven-day settlement to acquirers.

Read those two sentences as an operator and they say something the agentic half of the announcement does not.

The back end has already moved. While the industry argued about agent protocols and checkout buttons, the settlement leg quietly went onchain at a run rate that is no longer a pilot. Seven billion a year is not a demonstration. It is a line of business.

And then the asymmetry, which is the kind of detail that decides whether a treasurer sleeps. Issuers settle seven days a week. Acquirers do not yet. Which means, in the weeks before that changes, money can arrive on a Sunday and leave on a Tuesday, and somebody is financing the difference.

I have written variations of that sentence three times this year about three different products. The OCC's stablecoin proposal makes redemption take two business days while the token itself settles in seconds. Visa Direct's prefunding pilot exists because funding moves on banking days and payouts do not. And now the settlement leg itself is seven days on one side of the network and five on the other.

It is the same shape every time. The new rail runs continuously; the institutions attached to it do not; and the gap between them is paid for by whoever is holding the position when the calendar stops. If you want to know where the money is in this transition, it is there. It has been there the whole time.

Two registries

One more thing, and it is the one I would raise if I were on a scheme committee.

In December I wrote that Mastercard's acceptance framework created a new chokepoint nobody was naming: a registry of permitted agents, because whoever operates agent registration decides which agents are allowed to exist in the payment system.

Visa has now built one too. The Agentic Directory lists the agents and merchants Visa has verified as legitimate participants.

So there are two. They are operated by competitors, they will have different criteria, different onboarding, different revocation processes and different definitions of legitimate. An agent that is verified in one and unknown in the other is a different object depending on which network issued the card it is about to use.

That is a fragmentation problem the industry created in a single day, and it has a familiar solution — a jointly governed body, the way EMVCo was the answer to exactly this shape of problem forty years ago. Nobody has proposed one. I would bet against anyone doing so while both registries still look like a competitive asset, which is the same reason the protocols that touch money have not been donated either.

From a Panamá desk

Three consequences for institutions in this region, in increasing order of how soon they arrive.

The first is the registries, and it is the one with an owner problem. A processor here sits underneath both networks. Two agent directories with different criteria means two onboarding paths, two sets of evidence, two revocation regimes — and, downstream, an authorization decision that has to know which directory a given transaction's agent came through. Nobody in this region has been asked to prepare for that, and the preparation is not technical. It is deciding whose job it is.

The second is the weekend. Seven-day onchain settlement for issuing banks and not yet for acquirers is, in a dollarized economy, a straightforwardly good thing on one side of the house and nothing at all on the other. The issuing business gets its money on Sunday. The acquiring business does not. In a market where acquiring is the thinner and more price-pressured line, that asymmetry widens a gap that is already uncomfortable. Worth asking your network when it closes, and worth writing the answer down.

The third is the one I would not wait on. Nothing announced on 10 June has a general availability date. Not Intelligent Commerce, not Agent Score, not the directory, not the tokenized deposits layer, not Agent Pay for Machines. Which means, for a market our size, these arrive the way everything arrives: as a line in a processor release note, eighteen months from now, on a Tuesday.

The institutions that read the announcements now will have had a year and a half to form a position. The ones that wait for the release note will be implementing somebody else's.

My judgment

The interesting divergence this year is not Visa versus Mastercard on technology. It is Visa versus Mastercard on where the loss sits.

Visa is building the best possible decision. Score the agent, verify it in a directory, enrich the token data, run the request past a model trained on billions of transactions. Every piece of that makes the approve-or-decline more accurate, and none of it changes who eats the loss when the approval was correct and the outcome was still wrong — which, with agents, is going to be a large and novel category.

Mastercard is building an obligation. Not a well-specified one yet, but the word is on the page, and words on pages in this industry eventually become rules with effective dates and compliance obligations. If Mastercard follows through and publishes a rule, it will be the first party in agentic commerce to have done the hard half, and it will be able to charge for it.

My prediction, offered so it can be checked: within eighteen months one of these two will publish an actual rule — a defined liability allocation for agent-initiated transactions, with an effective date — and the other will follow within two quarters, because a network that cannot say who pays cannot compete with one that can. The protocols published by technology companies will not follow at all, for the reason I have given before: the ones closest to the money have the strongest reason to keep the question open.

And the third judgment, which is the one I keep arriving at from different directions. The agentic layer gets the announcements; the settlement calendar gets the money. Seven billion dollars a year moved onchain without a product launch, a keynote or a protocol war. Watch the back end.

Three questions for the people who own this

1. Two agent directories now exist, operated by competitors. Who in your institution is responsible for knowing which one a given transaction's agent is registered in — and if the answer is nobody, which of your systems would even carry that field?

2. Issuing banks settle onchain seven days a week; acquirers do not yet. For your institution, what is the current cost of that asymmetry in days of financing, and have you asked your network for a date?

3. Mastercard says settlement is guaranteed across multiple rails. Have you asked, in writing, what the guarantee covers, against whom it runs, and what happens when the rail is a chain your institution has never touched? The answer exists somewhere. It is not in the press release.

Reply to this email with a line or two. The sharpest answers will shape the next issue.

Sources

Visa, 10 June 2026 — "Visa Announces New AI, Stablecoin and Token Innovations to Power Intelligent, Programmable Commerce at Visa Payments Forum," issued from San Francisco. Source for Visa Intelligent Commerce; Agent Score and its development with New Generation; the Agentic Directory of verified agents and merchants; the strategic collaboration with OpenAI; the Large Transaction Model trained on billions of transactions; the Crypto Labs command line proof of concept described as an early development concept rather than a launched product; Tokenized Deposits and the statement that funds remain on banks' balance sheets; expanded stablecoin settlement pilots across regions, blockchains and currencies; the figure of approximately seven billion dollars annualized stablecoin settlement run rate as of March 2026; that issuing banks already settle onchain seven days a week and that Visa is working to extend seven-day settlement to acquirers; more than 160 stablecoin-linked card programmes live or in development; enriched token data and the token assurance signal; Pismo, Unified Checkout and Visa Intelligent Authorization; and the quotations from Jack Forestell, Chief Product and Strategy Officer, including "AI is transforming the front end of commerce. Stablecoins are reshaping the back end." The release gives no general availability dates for any product.

Mastercard, 10 June 2026 — "Mastercard Launches Agent Pay for Machines to Unlock Super-Fast, Always-On Payments," issued from Purchase, New York. Source for the description of Agent Pay for Machines as a service for high-volume, low-value, low-latency payments between AI agents and automated systems; that it extends the Agent Pay programme introduced in 2025; the four capabilities of credentialing with Verifiable Intent, permissioning with programmatically enforced rules and limits, transacting across providers and systems, and settling, where settlement is guaranteed across multiple rails; that settlement spans cards, accounts and stablecoins without a single named settlement asset; that some payments are only fractions of a cent; and the list of more than thirty initial participants and supporters including Adyen, Coinbase, Stripe, Cloudflare, Global Payments, OKX, Solana Foundation, Tempo and Utila. The release gives no general availability date, regional rollout, pricing, volumes or adoption metrics, and does not explain the technical mechanics behind credentialing or permissioning.

On what neither release says. Neither defines the liability allocation for an agent-initiated transaction that goes wrong. Mastercard does not define what its settlement guarantee covers, against whom it runs, or what happens when the rail is a third-party chain. Visa does not state that it guarantees anything about agentic transactions. Neither publishes registry criteria, onboarding requirements or revocation processes for its agent directory. Every question I raise about those points is a question, not a reported fact.

A note on what is mine. The reading that Visa's agentic announcements constitute an authorization posture while Mastercard's constitute a settlement posture; that sub-cent payments cannot clear on card economics and are therefore being routed to the partner chains named in Mastercard's list; that guaranteeing settlement across a third-party chain is a materially harder problem than guaranteeing it on your own rail; that the seven-day issuer and five-day acquirer asymmetry is financed by whoever holds the position across the calendar gap; and that two competing agent registries constitute a fragmentation problem with an EMVCo-shaped solution nobody has proposed — those are arguments, not findings. The prediction about a published liability rule within eighteen months is mine and is offered to be checked against the record. Nothing here reflects the position, portfolio, clients or vendors of my employer.