U.S. payments giant Mastercard announced on October 20 that it had opened early access for developers to Agent Pay for Machines (AP4M), a payment service for AI agents. The initial versions of buyer- and provider-side APIs, development tools, a testing sandbox, and documentation are now available, allowing developers to build and test commercial transactions between agents.
AI will do more than just “buy” things
The vision of AI agents researching, comparing options, and then making a purchase has been discussed for some time. But as it draws closer to reality, another question comes into focus: Who can use it, up to what amount, and with what authority? AP4M is designed to address precisely that question.
The service lets agents transact within authority and spending limits approved in advance by a company. Rather than handing over an unrestricted wallet, the company sets the boundaries first: “You can act within these limits.” It is much like delegating spending authority to an employee up to a certain amount.
Another feature is its ability to connect authenticated agents with service providers and support high-frequency transactions. This early access program offers developers a way to prototype transactions that incorporate these rules. The emphasis on setting boundaries first, rather than maximizing autonomy, feels like a characteristic move from a payments company.
Respond to a payment signal, then provide a credential
The system is built on x402, an open protocol for requesting payments online. The web has status codes that communicate what is happening, such as “404” when a page cannot be found. With x402, a service provider responds to a request made without payment with “402 Payment Required,” telling the other party that payment is needed. It is like a cashier telling a customer the price, translated into an exchange between machines.
In AP4M, the AI agent responds by creating a digitally signed payment credential that commits to payment within its approved limit. The provider verifies the credential and delivers the service, while multiple credentials can be consolidated for settlement later. The credential exchange itself, the company says, is handled off-chain.
This is useful for low-value, high-frequency transactions. Settling transactions worth just a few yen to a few dozen yen one by one would add to both the workload and the cost. Small payment commitments can be accumulated and settled together later. This approach, similar to batch settlement for point-card transactions, could be one practical way to handle payments between machines.
Demo shows nine payments from a single budget
The system was demonstrated at Cloudflare’s Cloudflare Connect event. An agent from marketing company Hype purchased research data and advertising images within a single approved budget, making a total of nine payments to two providers.
It is also worth looking at the setup behind the scenes. Coinbase’s wallet infrastructure was used to fund the agent, while the Base blockchain handled on-chain settlement. BVNK manages the agent’s wallet and, through the Mastercard Global Money Platform, processes payments and settlement in fiat currency or stablecoins.
The example shows an agent making different types of purchases from multiple providers using a single budget. It offers a glimpse of a process once handled by human employees—requesting quotes, routing internal approvals, and processing invoices—running automatically within a budget. The demo was limited to testing payments among multiple providers, but it concretely showed agents moving from “researching” to “procuring.”
A continuation of the initiative announced in June
Mastercard first announced AP4M on June 10 this year. At the time, it said the service would support settlement using multiple payment methods, including cards, bank accounts, and stablecoins. More than 30 companies and organizations were named as early participants and supporters, including Coinbase and Cloudflare, as well as Stripe, Tempo, Ripple, and the Solana Foundation.
AP4M is an expansion of the Agent Pay initiative Mastercard introduced in 2025. Building on Agent Pay, it broadens the scope to include high-frequency, low-value payments made continuously by AI agents and machines. The shift is from purchases that people review and pay for on a screen to transactions that machines keep paying for behind the scenes. The company that operates a card network is now moving to build the foundation for that shift.
In an era when agents can shop freely, what matters most may not be convenience itself, but how authority and budgets are designed. Who sets the limits, what can be processed automatically, and which transactions need to be verified? This early access program signals that developers can now start hands-on testing of that blueprint.
Comments