In one autumn, "agents that buy things" stopped being a demo and became infrastructure. Within a few weeks of each other, x402 went live at Coinbase, OpenAI and Stripe shipped the Agentic Commerce Protocol, and Google launched AP2 with sixty-plus partners. By March, Stripe and Tempo added the Machine Payments Protocol. A machine can now discover a service, agree a price, and settle, in a single HTTP round trip.
Read the stack straight and it separates into two floors. There is a commerce layer, where an agent discovers a merchant and checks out. And there is a rail underneath it that actually moves the money. Any commerce protocol can sit on any rail. The interesting question is which rails the layer above can reach.
The money is already moving
This is not a whiteboard market. By mid-2026 x402 has settled on the order of $50 million cumulative across 165 million transactions and roughly 69,000 agents on Base, at an annualized run-rate near $600 million. The average payment is about 32 cents, a size no card network can process at a profit. Visa, Mastercard, and Ripple have signed on to the standard.
The marketplaces to route it have appeared, and agents are already earning on them. One, Felix, reports over $260,000 from agent-run products. A sampler of what is live today:
And every serious forecast says today is the floor, not the ceiling. Morgan Stanley puts the agentic-commerce impact near $385 billion by 2030. Bain sizes the US market at $300 to $500 billion in the same window, and McKinsey has the US B2C opportunity alone approaching a trillion. The largest number is the machine-to-machine one: Gartner expects 90% of B2B buying to be agent-mediated by 2028, on the order of $15 trillion routed through agent exchanges. Against numbers like those, the $600 million on-chain today is a rounding error.
Two honest caveats, and both point the same way. By Artemis's read about half of that volume is still testing, not commerce. And nearly all of the rest is machines paying machines, for API calls, data, and compute, in fractions of a cent. It is a real new market. It is not the one where a person buys groceries, a merchant ships a box, and a tax authority wants a record. That economy still runs on Pix and cards. It has not moved on-chain, and it will not. The rail has to come to it.
The agentic economy has a settlement problem. It is not the one crypto thinks it is.
It was built in crypto, and in English
Look at where the energy went. The reference implementations, the SDKs, the popular directories, are by their own scope USDC on Base and a handful of other chains. The new agent marketplaces route stablecoins and virtual cards. It is excellent work, and it is almost all written from the United States and from crypto.
What it barely touches is the money the rest of the planet actually moves. Pix and account-to-account transfers. Boleto. Regulated cross-border FX. And the artifact a real purchase legally produces: a tax document, a mandate that says who authorized the spend, a receipt a dispute can be resolved against. In Latin America the interesting rail is not a card and not a coin. It is Pix, and an autonomous agent paying on it needs a scoped, capped, revocable mandate, not a bearer key it can leak.
x402 is not about crypto
Here is the move that made x402 work, with the coin taken out of it. The agent presents a pre-authorization, a signature that says "you may pull this exact amount, once." The facilitator relays it and settles inside the same HTTP request. The server verifies, the money moves, the resource is delivered, all in one round trip. Authorize once, settle in the request.
Nothing in that sentence requires a blockchain. It requires a pre-authorization the network can act on synchronously. Crypto happens to have a beautiful one. But fiat has one too, and it is older than crypto. A signed mandate is the fiat analog of that signature. A consumer signs, once, an authority with an amount cap, a merchant allowlist, and an expiry. Present it at pay time and it authorizes the debit in the same request, on Brazil's instant, government-auditable rail.
A crypto signature and a Pix mandate are the same primitive: one authorization, settled in band. The rail underneath is a detail.
So we built it, on both rails
An agent pays a live merchant over USDC via x402, or over Pix under a mandate, synchronously, through the same endpoint. Either way the payment closes with one hash-chained receipt that binds the mandate to the payment to the delivery. On the Pix side the delivery link is the NF-e, Brazil's government-validated invoice. The USDC leg is verified on Base; the Pix leg debits through the regulated rail with an end-to-end id the central bank can see.
The agent picks the one it holds. The merchant gets paid and gets a receipt. Nobody had to choose between "the future of money" and the rail their customers already use. The three ways to charge an agent, an API paywall, an MCP server, or a payment link, are public and copy-paste in x402-monetization-examples.
Where this goes
The near curve is easy to state. As agents get more capable, they consume more: more APIs, more data, more compute, more of each other's work. Every unit of that is a payment, thousands of tiny settlements per task, running in the background. In raw count, machine payments will dwarf human checkout. That curve is real, and today it is mostly crypto-native micropayments.
The volume that matters is the crossover. It arrives when agents stop only paying each other for API calls and start buying real things: a supplier invoice, a subscription, a payout to a person, a physical good. That money has to land on a regulated rail with a tax document. The incumbents already see it. Visa, Mastercard, and Ripple joined x402; AP2 shipped with Mastercard, PayPal, and Amex; Amazon and Stripe wired agent payments in. The fiat side is arriving, and it arrives regulated.
As it does, the mandate stops being a nice idea and becomes table stakes. A scoped, capped, revocable authorization is the only way a bank, a card network, or a central bank lets an agent move real money. Pix Automatico and Open Finance are already that shape in Brazil; stablecoin and agent-payment rules are catching up elsewhere. The rail-agnostic envelope, one authorization that settles as a signature or a mandate, is the natural endpoint.
And agents will not only spend. An agent that sells a skill through a paywall and reinvests the revenue in the data it needs is a business run by software, with a balance sheet. Payouts split between the agents that did the work. Every settlement leaves a verifiable receipt, and that receipt becomes the substrate for the next layer: agent reputation, dispute resolution, and one day financing an agent against its own auditable cash flow.
Latin America is the leading indicator. Pix is already the default an agent settles on, and the pattern generalizes: SPEI, UPI, instant rails everywhere. The layer that spans crypto and regulated fiat, with an evidence trail on every payment, is the one the real economy standardizes on.
The layer the real economy can use
The agentic-payments layer that wins will not be the one with the most USDC endpoints. It will be the one that treats a crypto signature and a fiat mandate as the same thing, and produces a receipt a buyer, a merchant, or a regulator can verify offline, without calling either party. That is the layer a bank can sit behind, a tax authority can accept, and a dispute can be resolved on. It is also, not by accident, the layer most of the world's payment volume runs through.
We are contributing the idea back where it belongs. Our proposal to the Basis Theory Agentic Commerce Consortium adds a Pix method, a mandate that doubles as the settlement authorization, and a Control Record to their Open Agentic Commerce spec, all additive and shipping with a working reference.
Machines learned to pay. Now teach them to pay the way the world already does.
Fabiano Cruz is co-founder of CodeSpar. See the full landscape, 236 companies, on the agentic-commerce map; read the OAC proposal; and the LatAm builder's index is awesome-agentic-commerce-latam.