Introduction

Until recently, software could read a webpage and suggest what to do next, but it couldn’t actually pay for anything. That’s beginning to change. AI agents now book travel, reorder inventory, query data providers, and buy compute for the businesses and people they work for. Once an agent can hold a budget and complete a payment on its own, a new kind of commerce opens up, and it brings real questions for the companies that move money. Visa worked with Artemis to look at the protocols, standards, and live onchain data behind agentic payments, and to get a clearer read on where this is heading.

Two kinds of agentic commerce

One of the report’s early points is that the term “agentic commerce” actually covers two fairly different things. The first is what you might call macro commerce. Here the agent is acting for a person, doing things like booking a flight or managing a subscription, and the payments look a lot like ordinary e-commerce. They carry consumer-sized values and sit comfortably on the card rails we already have. The second is micro commerce: small, frequent payments that one piece of software makes to another, often for an API call or a slice of compute. These usually come in well under a dollar.

Both are growing, but the micro side is where things get interesting. Payments that small don’t behave like human purchases, and that’s exactly where the economics of traditional payments begin to break down.

What the live data shows

Two open protocols went live over the past year to handle machine-native payments, and the report studies both. The first, x402, was incubated by Coinbase and Cloudflare and is now stewarded by the Linux Foundation. Since launching in May 2025, it has processed roughly $15.0 million in adjusted volume across 109.6 million transactions.¹ The second, the Machine Payments Protocol (MPP), was built by Stripe and Tempo with contributions from Visa. In its first few weeks after launching in mid-March 2026, it settled about $25,000 across roughly 115,000 transactions.

The two tell very different stories. x402 gives us a real look at sustained machine-native demand at scale, with most of its activity on Base, Solana, and Polygon.² MPP is much newer, but its two-party design and plans to support more than one settlement rail make it worth watching. On both, the average payment is a fraction of a cent. A fixed card fee on a transaction that small would cost far more than the payment itself.

Why small payments need new rails

None of this is a new ambition. Back in 1997, the engineers building the web set aside HTTP status code 402, labeled “Payment Required,” for a day when paying for content online would be as routine as loading a page. That day never came. Card fees were fixed, which made sub-dollar payments uneconomic, so the web ended up running on advertising instead.

What’s different now really comes down to two things happening at the same time. Capable AI agents have created steady demand to pay for resources at machine speed. And a newer generation of blockchains has pushed settlement costs down to fractions of a cent. Together, that makes payments in the one-cent-to-one-dollar range practical for the first time. Account abstraction, passkeys, and sponsored transactions help too, letting these rails run quietly in the background where users never have to deal with them.

Trust is the hardest part

Better rails only solve part of the problem. Traditional commerce assumes there’s a human doing the buying, someone with judgment who can be held responsible. An agent acting on delegated authority doesn’t fit that assumption. If an agent buys the wrong thing, or a malicious prompt redirects its spending, it isn’t obvious who should answer for it. Is it the person who handed over the task, the platform that ran the agent, the company that built the model, or the merchant? Existing legal and regulatory frameworks weren’t written with this kind of delegation in mind, and clear precedents may not be available yet.

Disputes get even messier. Chargeback windows and evidence rules were designed for human-speed commerce, where there’s usually one clear “order” to point to. Once agents are transacting thousands of times an hour, with money moving through chains of agents paying other agents, there’s no settled way to unwind a payment that went wrong.

The standards are starting to converge

A lot of standards are taking shape at once, covering everything from how agents communicate and prove who they are to how payments get authorized and settled. On the crypto-native side, that includes x402 and MPP. On the card side, protocols like the Trusted Agent Protocol, the Agent Payments Protocol, and Visa Intelligent Commerce build on decades of existing infrastructure for trust, fraud management, and authorization, and stretch it to cover transactions an agent starts.

The line between these two camps is already getting harder to draw. MPP now handles both onchain and fiat settlement through shared payment tokens, and several card-native protocols are adding stablecoin support. If anything, the two are starting to look less like rivals and more like parts of the same system.

What this means for payments

In all likelihood, this won’t come down to a choice between cards and stablecoins. Both will have a place. Cards are a good fit for the proxy and macro purchases that happen inside today’s merchant networks. Stablecoins suit the machine-native micropayments. And plenty of real flows will use both at different points in the same task. For payment networks, issuers, acquirers, and fintechs, agentic commerce is no longer a hypothetical. The practical question now is where along that range to focus.

That’s the direction Visa is working toward. The goal is infrastructure that doesn’t force the ecosystem to pick a side or rebuild on a separate stack: support for card-native trust and authorization, support for machine-native settlement, and a way across the two as they keep converging.