An Agent Cannot Open a Bank Account, So It Pays in Stablecoins

• An autonomous agent has no legal personhood and cannot pass a KYC check, so the banking system has no slot for it as a payer. A bearer instrument that moves without a human account holder is the only money it can hold on its own, which is the grounded reason stablecoins fit.
• The rails are real now. Coinbase’s x402 (May 2025) and Stripe and Tempo’s Machine Payments Protocol (18 March 2026) both let a server answer a request with a price and settle a sub-cent payment in seconds.
• Card rails and stablecoin rails are splitting the work. Visa and Mastercard agent credentials win the long-tail consumer merchants because of coverage, while stablecoins win agent-to-agent, API, compute, and data flows that card interchange was never built to price.
• The settlement asset will be whatever clears the regulatory perimeter. Mastercard’s machine framework settles only in regulated stablecoins, GENIUS is now US law, and MiCA already forced Tether off EU exchanges.
• US tax law treats stablecoins as property with no de minimis exemption, so every machine micropayment is a reportable disposal. The fix is mid-flight in Congress as of June 2026 and is converging on the GENIUS-compliant stablecoin from both ends.

An AI agent cannot open a bank account. It has no legal personhood, it cannot pass a Know Your Customer check, and it cannot hold a deposit in its own name, which Coinbase’s Brian Armstrong has described as a structural mismatch rather than a regulatory edge case to be smoothed over later. The conventional rails were built around a human account holder who can be identified, billed, and held liable, and an autonomous process running on a server is none of those things. So when the question becomes how a piece of software pays another piece of software for a unit of compute or a row of data, the banking system does not have a slot for the payer. A bearer instrument that settles in seconds without anyone needing to be the account holder does.

That is the case for stablecoins in machine payments, and it is a far more grounded case than the one the sector usually makes for itself. The pitch you hear at conferences is that stablecoins are the native currency of robots because they are programmable and global and cheap, which is true and also beside the point. The reason they fit is narrower and harder to argue with. An agent is not a customer of the financial system in any sense the system recognizes, and the only money it can hold and move on its own is money that does not require it to be a person.

What shipped in the last twelve months

The protocol race stopped being a thought experiment somewhere in 2025. Coinbase published x402 in May 2025, reviving the long-dormant HTTP 402 Payment Required status code so that a server can answer a request with a price, the agent constructs a USDC payment on Base or another chain, and the server settles before returning the resource. In its first six months x402 processed over 100 million payments, and on Base it settles a minimum payment of about a tenth of a cent in roughly two seconds. The x402 foundation now sits under the Linux Foundation with Coinbase and Cloudflare among its backers, and Cloudflare’s pay-per-crawl beta lets a developer fetch web data and settle the payment in the same request cycle.

The future of finance is here. Trade crypto and more

Stripe answered with the Machine Payments Protocol on 18 March 2026, co-authored with Tempo, the payments-focused Layer 1 it incubated with Paradigm. MPP runs over Tempo today but is built to be rail-agnostic, supporting stablecoins, cards, and buy-now-pay-later through Shared Payment Tokens, and Visa has already extended it to card rails on its own network. The early production users tell you what the rail is for, because they are not consumer storefronts. Browserbase lets agents spin up headless browsers and pay per session, PostalForm lets an agent pay to print and post physical mail, Prospect Butcher Co. lets agents order sandwiches for human pickup in New York, and Parallel charges agents per API call for web access. Stripe’s distribution did the rest, matching x402’s seller count within days of launch, which is the part that should worry anyone betting on a protocol rather than on a balance sheet.

Underneath the protocols, the wallet and identity layer is where most of the production work landed. Coinbase shipped Agentic Wallets in February 2026, smart-contract wallets with spending policies, session caps, and screening enforced before a transaction executes rather than reconciled after. Crossmint built lobster.cash with Visa, Solana, Circle, and Stytch, an open standard that hands an agent a virtual card and a USDC wallet in the same flow, with one set of spending controls governing both and the agent never touching raw card data or private keys. And at the standards level, ERC-8004 has emerged as the on-chain agent-identity standard co-authored by MetaMask, Google, the Ethereum Foundation, and Coinbase, which is the closest thing the space has to agreement on how an agent proves who it is acting for.

How an agent proves whose money it is spending

The harder problem than moving the money is establishing that the agent had the right to move it. With no human clicking a confirm button, the authorization has to travel with the request, which is why Google’s Agent Payments Protocol built the idea of a signed mandate, a cryptographic credential that records that a named user granted this agent the right to spend within a defined scope, and why the card rails and Stripe’s toolkit all consume those mandates as inputs while x402 treats the wallet signature itself as the mandate. Verifiable intent, in Mastercard’s phrasing, becomes the thing a dispute hangs on when there was never a person in the loop to point back to.

That authorization only means anything if the key it depends on cannot be lifted off the device, and since an agent has no human to re-authenticate each payment, its signing key has to live in a trusted execution environment, sealed away from the model and the application code, with spending limits and recipient allowlists enforced by the wallet contract before anything reaches the chain. Crossmint and Coinbase both run a two-signer version of this, an owner key held by the human or the platform and an agent key sealed in the enclave, and the design exists because a stablecoin is a bearer instrument whose holder owns it outright, so a compromized agent with an extractable key and a balance is a self-executing theft with no bank to call and no chargeback to file.

Where card rails win, and where they cannot follow

The temptation is to read all of this as a stablecoin story, and the consumer-facing half of it is mostly not. When ChatGPT checks out an Etsy order through OpenAI and Stripe’s Instant Checkout, or when a Visa or Mastercard agent buys something at one of the roughly 150 million merchants already on those networks, the settlement underneath is a card transaction tethered to a human account, dressed in a token so the agent never sees the card number. Visa unveiled its Trusted Agent Protocol on 14 October 2025 and Mastercard opened Agent Pay to its US cardholders the following month, and the appeal of both is coverage. An agent with a Visa credential can spend at any existing merchant without that merchant changing a line of code, which no stablecoin rail can match today. The cost of that reach is that the agent is permanently a delegate of a verified human, with the human’s account and chargeback rights still doing the real work.

Stablecoin rails win the half of the market the card networks were never built to serve, which is agents transacting with each other and with machines. A sub-cent payment for a single API call, a fraction of a cent for a data query, thousands of them an hour across jurisdictions, settling at all hours without a human in the loop, is a workload that card interchange and correspondent banking cannot price. This is the half where the native-currency claim holds, because there is no human account on either side to anchor a card transaction and nothing for a bank rail to attach to. A working shopping agent in 2026 therefore tends to carry both, a card rail for long-tail merchants and a stablecoin rail for API calls, compute, and agent-to-agent settlement, and the firms selling infrastructure have stopped pretending one displaces the other.

Visa, Mastercard, Google, Stripe, and Coinbase are each pushing a rival specification for how an agent identifies itself and pays, which one analyst quoted by American Banker called a land rush that leaves merchants confused and waiting to see which ecosystem reaches a critical mass of buyers before they commit. The way competing standards stalled the IoT build-out a decade ago is not an encouraging precedent for how fast that resolves.

The word that gates everything is regulated

Whatever becomes the settlement asset for machines will not be whatever is cheapest to move on-chain, because the institutions building the rails have already decided otherwise. When Mastercard expanded its agent payment work into Agent Pay for machines with more than thirty partners, it settled across six stablecoins, USDC, PYUSD, USDG, USDP, RLUSD, and SoFiUSD, every one of them a regulated coin, and it kept repeating the word regulated until the signal was unmistakable. Only stablecoins inside a recognized framework get to touch core payments infrastructure. The decentralization romance in most writing on this subject runs straight into that wall.

The GENIUS Act became US law on 18 July 2025, creating a federal licence for payment stablecoin issuers, mandating one-to-one reserves in cash and short-dated Treasuries, and banning issuers from paying interest to holders. In Europe, MiCA’s stablecoin provisions applied from 30 June 2024 and pushed the major exchanges to delist Tether for European users by the end of that year, since Tether did not seek authorization. The market that resulted is concentrated and legible, with dollar-pegged stablecoins sitting around 230 billion dollars by early 2026, Tether near 142 billion and Circle’s USDC near 60 billion, which is the supply the agent economy will draw on.

In crypto, one bad click is final. Learn to verify before you lose.

Mastercard’s machine framework introduced Know Your Agent, treating an agent as a prospective customer whose identity, scope of authority, and source of authorization have to be verified, with an evidence layer for dispute resolution where no human was in the loop. Crossmint holds crypto-asset service provider licences across all twenty-seven EU member states under MiCA and runs KYC, AML screening, and travel-rule compliance inside its wallet stack, because the alternative is asking every developer to become a money transmitter. The shape of the answer is clear even where the detail is not, and it looks like the agent inheriting a verified human’s compliance posture rather than acquiring one of its own.

Every payment is a taxable disposal

US tax law treats stablecoins as property which means every time a stablecoin is used, sold, or exchanged it is a disposal of a property asset and a reportable event, and there is no de minimis exemption under which a gain is too small to count. Whether the gain is a cent or a thousandth of one, it goes on Form 8949. Now picture the workload the entire sector is selling, an agent making thousands of sub-cent USDC payments an hour, and notice that each one is, on paper, a taxable disposal requiring cost-basis tracking. Stripe’s own guidance for businesses says it plainly, that no gain is too small under current law, which is a striking thing to read on the website of a company shipping a protocol whose pitch is frictionless machine payments.

The fix is mid-flight in Congress as of June 2026, and its shape says something about where this is going. The PARITY Act, reintroduced on 19 May 2026 by a bipartisan group on the Ways and Means Committee, would recognize no gain or loss on a regulated payment stablecoin unless the holder’s basis had fallen below 99% of its redemption value, which is a roundabout way of treating a compliant stablecoin as cash for tax purposes. A separate Senate bill from Cynthia Lummis proposes a broader 300-dollar per-transaction de minimis with an annual cap, and the House Ways and Means Committee held a discussion hearing on digital-asset taxation on 9 June 2026 with written submissions due the same week this is being written. Nothing is enacted. The relief that looks most likely to pass is the one narrowed to regulated payment stablecoins, which means the tax fix and the machine-payment use case are converging on exactly the same instrument, the GENIUS-compliant stablecoin, from opposite ends of Washington.

Until something passes, the honest read is that the US tax treatment makes high-frequency machine micropayments irrational for any payer inside its reach, in the same way it has kept people from spending Bitcoin at a coffee shop for a decade, because the settlement takes two seconds while the reporting obligation behind it does not shrink with the size of the payment.

Pay per call is real, the treasury story less so

Parallel was built on the premise that agents are the primary users of the web and should pay per API call rather than holding a seat in a monthly plan, and the same logic runs through pay-per-session browsers and pay-per-query data. An agent optimising cost per action has no loyalty to a flat fee and no reason to pre-commit to capacity it might not use, so usage billing suits a machine buyer in a way it never quite suited human ones, who mostly prefer the predictability of a subscription even when it costs more.

The more excitable claim, that fleets of agents will park idle stablecoin balances in DeFi and run themselves as little treasuries earning yield, deserves more scepticism than it usually gets. GENIUS bans the issuer from paying interest on the coin itself, so any yield has to come from lending or liquidity provision, which reintroduces precisely the smart-contract and counterparty risk the stablecoin was chosen to avoid, and the yield is taxable income at the moment it is received. A treasury that generates a reportable income event every time it earns a fraction of a cent is not obviously a feature. The one place machines already earn and spend at scale is the decentralized physical infrastructure networks, where Helium and its successors pay devices for providing coverage and let them spend those earnings on usage, and that model works because the device is a node in a market.

What the live numbers say

x402’s hundred million payments in six months is real momentum, but at a tenth of a cent of minimum value the dollar volume is small, and one count put the addressable agentic-payment market between February and March 2026 in the low tens of millions of dollars. The sellers number in the low thousands and the consumer flows that generate headlines are running on card rails that have nothing to do with stablecoins. This is early infrastructure being described in the language of a finished economy.

For software agents transacting with each other and with services, stablecoins are becoming the settlement layer, because the banking system has no place for a payer that is not a person and card interchange cannot reach a sub-cent payment made a thousand times an hour. For the consumer who wants an agent to buy something from a normal merchant, the rail is a card with a token wrapped around it and a human still on the hook. The winners look less like a token and more like the firms that own distribution and compliance, Stripe and Coinbase and the wallet layer between them, and the binding constraint over the next two years is not throughput or latency but whether a US tax provision and a regulatory definition of a permitted coin land in a place that lets the machine economy run the way its builders keep describing it.