AI agents are emerging as a new class of crypto users, making machine-to-machine payments in stablecoins like USDC. Coinbase, Cloudflare, Visa, and Mastercard are racing to build the rails for this new “agentic economy.”
For years, crypto companies have talked about onboarding the “next billion users” into the industry. That wave may finally be arriving — just not in the form anyone expected. Rather than underbanked individuals or people fleeing hyperinflation, the industry’s next major user base might not be human at all. It might be software.
A New Class of Customer
AI agents — software capable of independently carrying out multi-step tasks — are increasingly transacting on their users’ behalf. To complete many of these tasks, an agent often needs to pay for something along the way: data, computing power, or access to an online tool. That simple need has quietly kicked off a genuine race among crypto and fintech companies to build the payment infrastructure these agents will rely on.
Coinbase has developed x402, a payment protocol built specifically for AI agents. MoonPay’s PayBox gives agents access to a user’s cards and crypto wallets. Cloudflare recently rolled out its own Wallets product and a companion tool called cloudflare.pay, designed to let AI agents make online purchases within preset spending limits while still letting sellers identify the human behind a given agent.
What currency these agents will actually use to pay remains an open question — and increasingly, it’s the real battleground. The major players in payments aren’t debating whether machines will drive the next wave of commerce. They’re competing over what form of money will power it.
Why Stablecoins Have the Early Lead
So far, stablecoins hold the strongest position in payments built from the ground up for software, while traditional card networks remain dominant for larger purchases where buyers expect credit access, refunds, and dispute protection.
Stablecoin issuer Circle is currently testing USDC micropayments, and even BitMEX co-founder Arthur Hayes has said he’s coming out of retirement specifically to build a currency for this emerging “agentic economy.” Coinbase said earlier this year that its x402 protocol had already processed more than 165 million payments worth a combined $50 million.
The mechanics are straightforward: an agent receives a price, sends payment, and receives what it requested — all without creating an account or manually entering card details, and without a human involved beyond the original instruction that triggered the purchase.
Lincoln Murr, Coinbase’s head of AI product, told CoinDesk he estimates roughly 99% of these x402 payments currently use USDC — an early but notable lead for dollar-denominated stablecoins in the one corner of this market where usage can actually be measured.
What Agents Are Actually Buying
Contrary to the popular image of AI agents booking vacations or shopping for shoes, the current reality is far more mundane. Most of these purchases are agents paying small amounts — often just cents — for data, computing power, or access to APIs, the connections that let one piece of software request information or services from another.
“The primary product-market fit today is machine-to-machine payments from agents to APIs,” Murr explained. “Given the velocity of these AI agents and how quickly they move across the internet, we need a standard way for them to pay for things.”
In raw dollar terms, the activity is still tiny. x402 processed roughly $24 million over 30 days in July — a volume Visa handles in about one minute. Murr compares the current moment to the early, chaotic days of internet file-sharing: “We’re still in the ‘Napster/LimeWire era’ of agentic payments… If we think about agents as the evolution of apps, we’re in the ‘QR code scanner/flashlight app’ era. People are experimenting, and we know it’s the future, but we’re still figuring out the exact mechanics.”
Even so, the underlying structure hints at why payment companies are paying close attention. A single software agent completing one task might end up buying from dozens of different services along the way, effectively replacing a bundle of monthly subscriptions with a continuous stream of tiny, individual purchases.
How a Machine Actually Pays Another Machine
Coinbase’s x402 protocol offers a useful window into how this works in practice. It functions essentially as a paywall built for software: when an agent requests data or a service, the seller responds with a price. The agent pays, and once that payment is verified, the results are released — no subscriptions, no manual checkout process.
The name comes from an old, rarely used web status code called “402 Payment Required,” which Coinbase repurposed into a functioning standard that lets online services request payment in a format software can understand and act on automatically.
This shifts the underlying economic model in a meaningful way. Where subscriptions dominate today, agents could instead pay on a strict need-to-use basis. A human would likely balk at paying a fraction of a cent for a single piece of data, but an AI agent has no such friction or hesitation.
Coinbase’s most recent update on the x402 ecosystem counted more than 480,000 active agents, with an average transaction size of roughly 30 cents. Most of that activity happens on Base, Coinbase’s own layer-two blockchain. Murr estimated that 25% to 30% of those transactions may actually be coming from users testing the system or trying to climb public leaderboards, rather than genuine service purchases — a reminder that current usage numbers likely overstate real organic demand.
Why Crypto Fits This Use Case So Well
The appeal of stablecoins for this specific use case comes down to simple economics. Murr put typical card acceptance costs at 2% to 4% — fees that become genuinely awkward once the item being purchased costs less than a dollar, especially once fixed processing costs are factored in.
Stablecoins, by contrast, can move around the clock, and a merchant or payment processor can cover the underlying blockchain fee directly, meaning an agent doesn’t necessarily need to hold a separate cryptocurrency just to complete a transaction.
“Stablecoins are particularly well suited to the kinds of transactions we expect AI agents to make at first, essentially lots of very small, high-frequency payments for things like API calls, data, inference, and content,” said Stephanie Cohen, Cloudflare’s chief strategy officer.
Cloudflare has designed its Wallets product to function more like a corporate card than an open-ended crypto wallet. An operator deposits funds into a main wallet, then allocates a smaller allowance to an individual agent, along with rules governing its budget, approved sellers, and maximum purchase size. “Those guardrails are really what make that autonomy possible,” Cohen said. Within those defined limits, the agent can spend freely without requiring approval for every single transaction.
Cloudflare’s system can also batch numerous small purchases together before recording a single combined payment on the blockchain, reducing fees, and pair that with escrow arrangements that hold funds until a seller actually delivers what was purchased. The company’s Monetization Gateway, still in early stages, would eventually let websites charge for individual pages, datasets, or tools rather than requiring a full subscription.
Much of this infrastructure isn’t fully live yet. Users can currently claim an identity through cloudflare.pay, but Cohen says full funding, withdrawals, and agent wallets are still “expected in the coming months.” Circle is testing a similar rapid-confirmation model through its Nanopayments product, while MoonPay’s PayBox — launched July 29 — connects directly to AI assistants like Claude or ChatGPT and gives them access to both cards and crypto wallets, letting a single agent handle everything from travel bookings to crypto payments through one interface.
Cards Aren’t Sitting This One Out
Despite stablecoins’ early lead, traditional card networks aren’t conceding the space. Mastercard’s approach centers on what it calls a digital spending voucher: an owner defines what an agent is allowed to buy and how much it can spend, a seller verifies those rules, delivers the service, and claims payment afterward.
“Execution can happen continuously in real time, while settlement follows a more efficient batched model,” said Sapan Mandloi, Mastercard’s executive vice president of tokenization and checkout services. Sellers under this system can choose to receive payment in either fiat currency or an accepted stablecoin, meaning buyer and seller don’t need to transact in the same form of money at all.
“Our view is that agentic commerce will be a multi-rail environment,” Mandloi said. “We see stablecoins as complementary to existing payment systems, not a replacement for them.” Cards retain one major structural advantage: they’re already accepted at millions of businesses worldwide, backed by established credit access, dispute resolution, and familiar user experiences that stablecoins currently lack for larger transactions.
Banks are testing similar concepts on existing infrastructure. In February, DBS and Visa demonstrated an AI agent purchasing food and drinks using standard DBS/POSB credit and debit cards, with DBS group head Ananya Sen saying the trial proved agent-led payments could be deployed “securely and safely at scale.” The two companies are now exploring extending the concept into online shopping and travel bookings — larger purchases that, unlike a 30-cent API call, benefit from established refund and dispute systems already built around card networks.
What Happens When the Agent Gets It Wrong
Moving money securely is only part of the challenge. An AI agent can misunderstand an instruction, select the wrong service entirely, or follow a malicious prompt even when the underlying payment mechanism itself functions exactly as designed — meaning control over the software’s behavior matters just as much as the payment rail carrying the money.
Coinbase, Cloudflare, Mastercard, MoonPay, and wallet infrastructure company Turnkey are all converging on a similar first line of defense: restrict what an agent can do before it acts, using combinations of capped balances, pre-approved sellers, transaction limits, and human approval requirements for particularly sensitive actions.
“Agents need to be granted scoped permissions on a wallet to prevent damages,” said Turnkey co-founder and CEO Bryce Ferguson. “You can almost think of this like a self-driving car. In the early days, self-driving cars needed somebody with their hands on the steering wheel to make sure nothing went wrong. Right now, we’re in that phase for agents, and this will evolve over time where we can trust agents to have more and more scope and permissions to take actions without a human in the loop.”
Ferguson also drew a clear distinction between markets genuinely built for software and the broader consumer economy: “Autonomous agents don’t need crypto wallets, strictly speaking. They can use credit cards and bank credentials, but that will likely be for old-school purchases like a pair of shoes or groceries. Crypto is ultimately a much better rail for purchasing a single API call or a single piece of data.”
Transaction size also changes how much fraud protection actually makes economic sense. “For micropayments, mitigating fraud can often cost more than the value of the fraud,” Cohen noted, meaning a seller’s reputation may matter more than a full dispute process at small scale — while “for larger payments, card rails remain a better fit given fraud and other protections.”
Coinbase’s Murr echoed the same logic: “If my agent messes up and wastes 10 cents, I’m not going to litigate over that. However, as we move into larger purchases, we want to add support for escrows and refunds.” Coinbase’s developer documentation already outlines checkout flows where buyers approve USDC payments before a seller can collect them, with support for partial or full refunds on completed transactions. Still, one fundamental question remains genuinely unresolved across the industry: who bears responsibility when an agent follows all the rules correctly and still ends up making the wrong purchase.
Still Very Early Days
Stablecoins currently hold a real, measurable lead in the agentic payments race, and that lead could widen further if digital services increasingly shift toward per-request pricing instead of monthly subscriptions. But the broader market remains far behind even its own most optimistic predictions. Cloudflare’s wallet infrastructure is still in development, Mastercard’s solution remains in early access without published adoption numbers, and Turnkey says activity isn’t yet operating at meaningful scale. A significant portion of Coinbase’s own reported transaction volume may reflect testing and leaderboard-chasing rather than genuine commercial demand.
There’s also a more basic, practical hurdle standing in the way: simply getting money into an agent’s hands in the first place. “Getting set up with the wallet is still a massive pain point,” Murr acknowledged.
The eventual market will likely end up using several payment methods simultaneously — an agent presented with a single price and spending rule, with the human behind it choosing to pay by card, bank account, or stablecoin, while the seller decides how they’d prefer to receive funds. Crypto doesn’t need to fully displace cards to become the dominant payment method for machines specifically; it simply needs to keep reducing friction for the humans setting these agents up in the first place. “To make it frictionless, we’re enabling seamless setup guides,” Murr said. “Ultimately, you should just be able to tell your agent, ‘Set up a Coinbase agentic wallet for me,’ and it will go search the internet, download the wallet, and prompt you to sign in via email and fund it with crypto.” Funding that wallet in the first place, he added, remains the hardest part of the whole equation — one Coinbase is looking to solve with a familiar tool: fiat onramps.
Frequently Asked Questions
Q: What are AI agent payments? AI agent payments refer to autonomous transactions made by AI software on behalf of a user, typically to purchase small digital services like API access, computing power, or data, without requiring a human to manually approve each individual purchase.
Q: Why are stablecoins currently favored over credit cards for these payments? Stablecoins avoid the 2-4% acceptance fees typical of card networks, which become impractical for purchases costing only a few cents. They can also move around the clock, and merchants can cover blockchain fees directly, removing friction for the software making the purchase.
Q: What is Coinbase’s x402 protocol? x402 is a payment protocol that functions like a paywall for software, allowing an AI agent to receive a price for a service, send payment, and receive access, all automatically and without a human manually creating an account or entering payment details.
Q: How much money is actually moving through agentic payments right now? Volumes remain small in absolute terms. Coinbase’s x402 processed roughly $24 million over 30 days in July, an amount Visa’s network handles in about one minute, reflecting how early-stage this market currently is.
Q: What happens if an AI agent makes a mistake and buys the wrong thing? Companies are still working out the answer. Current safeguards focus on limiting what an agent can do beforehand through capped balances and approved sellers, but responsibility for a technically successful payment that fulfills the wrong request remains an unresolved question across the industry.
