Most conversations about AI and payments focus on agents shopping for people — ordering groceries, booking flights, comparing prices. But there’s a less visible and larger shift happening underneath: AI agents paying other AI agents directly. Not on behalf of a consumer. As part of their own workflows. When one agent needs data, compute or another agent’s capabilities, it pays for access in milliseconds. No human involved on either side.
What are ‘machine payments,’ exactly?
There’s an important distinction between agentic commerce and machine payments. In agentic commerce, a human is always the principal — the agent acts as a delegate, buying something a person wants. In machine payments, the transaction may be entirely between software systems pursuing goals that no human initiated in the moment.
The examples are concrete and multiplying. An AI agent spins up additional cloud compute when it hits a bottleneck and pays per inference. A research agent purchases a real-time data feed to complete an analysis. One model pays another for a specialized capability it lacks. API services price per call and settle automatically, with no invoicing cycle and no procurement department.
What makes this possible is that AI agents have none of the friction that has historically made small payments impractical. They have no attention limits, no tolerance for multi-step onboarding, no hesitation around spending two cents. The transacting entity is software — running autonomously, operating around the clock — and the entire payment architecture needs to match that. Agents are the customer type that micropayment infrastructure was always designed for. They just weren’t the customer anyone expected.
Why existing payment rails don’t fit
Traditional banking was built for humans. Account creation assumes identity documents. Authentication assumes physical presence. Settlement takes hours or days. And the minimum transaction economics of card networks — roughly thirty cents plus 2.9 percent per transaction — make a two-cent payment structurally impossible. The infrastructure and the use case are simply mismatched.
What agents need is instant settlement, fractional-cent transactions, programmatic authorization and the ability to establish payment relationships entirely in code.
What’s being built
Two protocols are leading the early infrastructure, representing complementary approaches.
Stripe’s Machine Payments Protocol, or MPP, is an open standard co-authored with Tempo that gives agents an internet-native way to pay. An agent requests a resource from any HTTP endpoint. The service responds with a payment request. The agent authorizes and receives the resource — all programmatically. MPP supports both stablecoin and fiat settlement, which makes it practical for businesses already running on traditional payment rails; the same Stripe API that handles human transactions can handle machine ones. The design goal is programmable spend authorization: not just access to a resource, but a verifiable, scoped permission to pay for it.
The x402 protocol, developed by Coinbase as an open standard, takes a different approach: it revives the HTTP 402 “Payment Required” status code that has sat dormant in the web’s architecture since the 1990s. An agent requests a resource, receives a 402 response with payment instructions, pays instantly in stablecoins and gains access — one line of code on the server side. According to Coinbase, the protocol has already processed over 50 million transactions. Where MPP is optimized for businesses that want fiat compatibility and Stripe’s existing settlement infrastructure, x402 is optimized for stablecoin-native deployments that need the lightest possible integration path.
To make this concrete: imagine a research agent that needs to read a paywalled New York Times article to complete its task. Today, that’s a dead end — no account, no subscription, no way to pay. With x402, the Times publishes a per-article price to its API endpoint. The agent hits the endpoint, receives a 402 response with a stablecoin payment address and a price of, say, $0.25, pays in milliseconds, and receives the article. No login, no subscription tier, no checkout flow. The transaction settles in fractions of a cent of fees. For publishers that have spent a decade wrestling with paywalls and ad revenue erosion, this is a different kind of business model — one where the reader might be software.
Stablecoins are emerging as the natural settlement layer for both protocols. They’re programmable, instant, fractional, borderless and operate continuously — matching how software works, not how banks do. In 2025, stablecoin transaction volume reached roughly $33 trillion, rivaling Visa and Mastercard combined.
Why this matters beyond tech
Machine payments unlock business models that weren’t previously viable at scale. Content priced per article rather than per subscription. APIs charged per call rather than per tier. AI services billed per task rather than per seat. The practical implication is that every API endpoint becomes a potential product with real-time pricing — and the buyer doesn’t need to be a human with a credit card and a billing department. A research firm could expose its proprietary dataset to agents at $0.001 per query and run a meaningful revenue line without a single sales conversation. A specialized AI model could charge other models for its capabilities on every invocation. The economics of building and selling software change when the purchasing friction drops to nearly zero.
The execution question
The protocols define how machine payments are initiated and authorized. What they don’t solve is what happens when something goes wrong — and in fully automated, high-frequency agent workflows, things will go wrong. An agent that misfires a payment with no human in the loop, no dispute path and no audit trail is a different category of problem than a failed human transaction. The infrastructure needs to guarantee that a failed mid-sequence payment doesn’t result in double charges or resources delivered without authorization; that every transaction is logged in a form that supports accountability; and that agents operating across multiple systems can manage payment credentials without exposing them. That execution layer — the substrate that makes machine payments reliable enough to trust at scale — is still being built. The protocols are ahead of it.
Where this is heading
Machine payments aren’t a future trend being modeled on spreadsheets. The protocols exist. Transactions are happening at scale. What’s still missing is recognition that this represents a genuinely new category of economic actor — not a faster version of human payments, but something structurally different. When the majority of internet transactions don’t involve a human on either side, the frameworks for trust, reliability and accountability have to be rebuilt from the ground up. The companies and developers who understand that now, while the architecture is still being designed, will have shaped it. Everyone else will inherit whatever they built.
This article is published as part of the Foundry Expert Contributor Network.
Want to join?
Read More from This Article: Machine payments: When agents start paying agents
Source: News