At this year’s Jackson Hole symposium, Princeton economist Markus Brunnermeier warned central bankers that AI agents operating at scale could outmaneuver regulators and make markets more erratic. He argued that central bank transparency “arms the opponent,” which is to say the market, and particularly once artificial intelligence (AI) can predict policy. He suggested two press conferences: one for humans, and one for machines, designed as training data. The problem: AI can listen to both.
Policymakers have been reckoning with how AI affects payments, tokenization, monetary policy, and financial stability ever since it morphed into acting. Back in August, a US appeals court lifted an injunction that had barred Perplexity’s shopping agent from Amazon, ruling that when an agent acts on a user’s instruction, it is the user who accesses the platform. Amazon has since blocked Muse, Meta’s personal AI agent, in a similar dispute over how autonomous software interacts with existing platforms. Elsewhere, the White House has called for autonomous laboratories and AI-run research. And four of Google’s most senior AI researchers, led by Jeff Dean, have left to build a company that runs experiments at machine scale. Now that agents have moved to carrying out tasks, they will increasingly need to access data, buy services, and transact with other systems. How software pays is moving from theory into practice.
Still, what money these agents use—and what that means for the dollar—is not yet settled. That choice is being made now, by market forces and policymakers, through the payment rails being built and selected for machine transactions and the rules that govern them.
What changes when machines transact
Payments built for people assume a person is in the loop: an account opened with identity checks, a human to approve or dispute a charge, settlement in hours or days, and fees a buyer will tolerate a few times a day. Agents break all of these assumptions. They act on delegated authority, spending within limits a person sets and changes at will. They can transact around the clock, whether or not the user is online, at fractions of a cent whenever they use data or compute. They also turn familiar questions into system-design problems: how do you prove which agent is acting and for whom? How do you authorize it, limit its standing privileges, and decide who is liable when it makes a mistake? The Perplexity case shows how unsettled these questions still are.
The rules are being forged now. Google’s Agent Payments Protocol (AP2), for example, records a user’s instructions in cryptographically signed mandates that bind an agent to a specific purchase, giving merchants proof that the user authorized it. Coinbase has built x402 into AP2 to let agents pay in stablecoins directly over the web, including very small payments. Visa (Trusted Agent Protocol) and Mastercard (Agent Pay) are likewise developing approaches to agent identity, authorization, and payment.
Which rails win is still open
Digital, programmable money is blurring the line between currency and financial assets. Tokenized deposits and Treasury funds can move, settle, and serve as collateral on the same rails. However, under the GENIUS Act, the US stablecoin law enacted in July 2025, payment stablecoin issuers cannot pay holders’ yield. That makes programmable money and blockchain-based smart contracts a natural fit for machine commerce, but it doesn’t reveal which form of money agents will use. Several options are competing: dollar stablecoins, tokenized bank deposits, agent-payment systems that card networks are building, and, in some countries, central bank digital currencies. The outcome is not predetermined. Regulation, market validation, and trust will shape it as much as technology.
Stablecoins have an early structural advantage. Around 97 percent of stablecoin issuance is dollar-denominated, and the digital currency offers continuous settlement and the ability to handle small-value transactions at low cost, features well suited to machine-to-machine payments. That makes stablecoins an early contender, though not necessarily the eventual winner, for the money of the machine economy.
What it could mean for the dollar
If dollar stablecoins become a dominant payment rail, they could strengthen the dollar’s global position. Widespread use of dollar-denominated stablecoins could extend the dollar’s reach into the machine economy and even become a tool of economic statecraft. Federal law requires permitted stablecoins to be fully backed by cash, short-term Treasuries, and similar liquid assets, and issuers already hold a large share of their reserves in US government debt. At the end of 2025, Tether alone reported about $122 billion in direct holdings of US Treasuries ($141 billion including repurchase agreements). That means the company holds more US debt than the entire state of Germany does. The stablecoin market is now more than $300 billion; Citi projects it could reach $1.9 trillion by 2030, potentially creating a substantial and largely foreign stream of demand for US government debt while extending the dollar’s network effects into machine commerce.
Some analysts, among them the macro strategist Luke Gromen, argue that a reliable base of stablecoin demand for Treasuries could let Washington finance part of its debt more cheaply, for example by steering reserves toward lower-yielding securities. At scale, that would amount to financial repression, holding yields below market levels at a cost to savers. But that lever only works once stablecoin reserves are large enough to move Treasury yields. At their current size, they are not there yet.
US fiscal pressures make the new structural demand for Treasuries all the more important. US federal debt is now above $40 trillion, and the country has lost its top credit rating from all three major agencies. If stablecoins generate sustained demand for Treasuries, they could strengthen the dollar’s position.
The risks, and what policymakers should do
A handful of stablecoin issuers dominate this market. Tether and Circle together hold about 85 percent of the entire market. As they now in aggregate rank among the most significant holders of US government debt, a loss of confidence in a large stablecoin could trigger rapid redemptions and forced sales of Treasuries, transmitting stress into broader markets.
The speed of machine-driven transactions adds another vulnerability: AI agents can transact continuously and at scale, while parts of the financial system still rely on settlement and risk-management processes designed around human trading hours. In other words, an agent-induced run could build while an issuer is unable to sell reserve assets because markets are closed.
Those risks are manageable, and rules to do so are being written now. Policymakers face two tasks:
1. Finish the US rulebook with these risks in view. The Office of the Comptroller of the Currency proposed its GENIUS Act rules in February, the Federal Deposit Insurance Corporation in April, and Treasury in August, and the law takes effect by January 18, 2027 at the latest. The Act already requires full backing, disclosure, and redemption. The final rules should add four things:
- closer supervision for the issuers that hold most of the market;
- a cap on how much of any issuer’s reserves can sit at a single bank;
- reserve reporting frequent enough to keep up with machine-speed redemptions;
- a wind-down plan that can start before a failing issuer has to sell bills into a stressed market.
2. Coordinate standards with international partners. Industry is already writing the technical rules for agent payments through AP2, x402, and the card networks’ protocols. Governments now must decide how those payment rails and forms of money will work across borders. The GENIUS Act already gives Treasury one lever: it can recognize a foreign stablecoin regime as comparable to the US one, which determines whether an international issuer can serve US users. Using that lever with close partners and aligning rules on agent identity through the BIS and the Financial Stability Board could help dollar rails operate across jurisdictions.
If that coordination works, and AI agents come to rely on dollar-denominated assets, stablecoins, and US payment infrastructure to transact, the dollar would become embedded in the operating system of machine commerce. That makes the rules being written today consequential for US financial power well beyond payments.
Ivan Ferrari is a nonresident senior fellow at the Atlantic Council’s GeoTech Center.
The GeoTech Center champions positive paths forward that societies can pursue to ensure new technologies and data empower people, prosperity, and peace.
Further reading
Image: Arturo Anez via Unsplash