AI agents will buy and sell on our behalf, yet today’s payment systems still assume a human sitting in front of a screen clicking "confirm."

In early 2015, while at Meta, our company acquired Wit.ai, a natural language startup. We began collaborating with Facebook AI Research (FAIR). That summer, we tested M, a mobile assistant in Messenger—designed to chat like a friend. You’d ask it to book a restaurant, send flowers to a loved one, or plan a trip—it would execute everything seamlessly, without touching any app. Users quickly fell in love.

When launching, I described M as “artificial intelligence trained and supervised by humans.” That was true—but we never managed to automate more than 30% of conversations. At the time, large models hadn’t emerged; model capabilities were insufficient. The result: demand outpaced supply, and the product ultimately had to be shut down.

In 2019, we launched Libra—a digital asset native to the internet, flowing globally like a message. Libra died due to political headwinds; that story has been told elsewhere. I remain grateful for Mark Zuckerberg’s passionate support of both ventures.

I misjudged the timing by several years both times. M required models capable of autonomous action; Libra required governments to recognize digital assets. Now, both have arrived. Models are good enough; stablecoins—essentially prototypes of Libra—are now enshrined in law across the U.S. and Europe.

The two bets I made too early are now converging simultaneously. This deserves serious attention. In my view, together they signal the death of apps and screen buttons—and the emergence of an entirely new paradigm for capital flow.

The Interface We Know Is Disappearing

By "interface disappearing," I mean the familiar UIs—apps, websites, buttons—all vanish. Of course you’ll still view photos and maps; I’m not that extreme. How we interact with AI and computers will become indistinguishable from how we interact with people: speak, type, and tasks get done.

For over 40 years, we’ve learned computer language: programming, menus, forms, dropdowns, hamburger icons, “Are you sure?” popups. We’ve become so accustomed to it that we’ve forgotten we’re constantly translating. Now, computers are speaking our language.

This also redefines what an “operating system” means. Today’s OS is primarily an app launcher—those who control the launcher set all downstream rules. When people no longer launch apps, the OS becomes the agent you converse with—and right now, no one owns it.

Where attention concentrates will shift. For the past two decades, human attention has been focused on information feeds and apps. Going forward, it will concentrate on agents: personal agents working for individuals, enterprise agents serving companies. They’ll increasingly make decisions on our behalf.

Whoever controls the winning agent will become the most powerful company of this era. I see no room for debate. Meta’s Muse is doing exceptionally well—and only just beginning. Google will almost certainly enter. Frontier model companies won’t be absent either. Any firm capable of building strong economic network effects among stakeholders will compete fiercely. Whether these agents succeed, and whether companies profit, hinges entirely on business—because inference, virtual machines, and foundational model training all cost money. Moreover, an agent that can’t spend isn’t truly capable.

Payments Are Built on Buttons

At PayPal, the product was the button. During my years leading PayPal, we spent years optimizing the button experience—cutting steps, shaving seconds—to help more people complete checkout. Now, I’m writing its obituary.

Today, nearly every financial product is built around a screen: banking apps, checkout pages, card number forms, SMS verification codes sent to your phone. The entire risk framework of payments—authentication, fraud detection, dispute resolution, liability—assumes a human in many stages. Remove that person, and the entire assumption collapses.

When your agent buys from a corporate agent, there’s no checkout page, no one to send a verification code. The question becomes: Do you authorize this agent to make such payments? This question wasn’t needed when the current architecture was designed. Card networks and emerging protocols are now seeking answers—the rest must follow.

The interface for digital funds will undergo profound change because money will no longer flow through UIs.

Stablecoins Are No Longer the Answer

It may sound odd coming from a Libra co-founder: I believe stablecoins are no longer standalone entities.

This framing has been useful. It gives regulators something concrete to govern, and stablecoins have thus gained recognition as legitimate digital assets. Europe’s MiCA is fully implemented; the U.S. GENIUS Act is law and will take effect in January. After years of struggle, we’ve largely won this battle. Stablecoins equal acceptable digital assets.

But stablecoins alone cannot solve most real-world problems. Holding digital dollars is indeed useful. Using them effectively is often difficult. To be truly scalable, stablecoins must integrate with these components:

Agent-Native Payment Layer

Combine all of these: stablecoin settlement, account management, fiat connectivity, card issuance—but no frontend. Just a simple API layer that agents can use anytime they need to transfer value—whether buying or selling on behalf of individuals or corporations. This is the “agent-native payment layer.”

The API is simple; the backend is complex. You need licenses, banking partners, global card issuance and acquiring, real-time connections to payment systems in over 100 countries, liquidity, compliance—all operating simultaneously. This is the hardest, most grueling work in payments—and that’s why so few companies succeed.

You tell your agent: “Fix the fence by weekend.” It finds a contractor—more likely the contractor’s agent—negotiates price, pays deposit, then releases final payment upon receipt of photos showing completion. You never touched an app. Never went through a payment flow. You simply said what you wanted.

Consider a company whose agent manages a fund and a set of rules: pay 400 contractors across 30 countries every Friday, increase ad spending when data permits, settle cloud bills, issue refunds to customers with damaged orders. Each transaction lands in different endpoints: a bank account in one country, a mobile wallet in another, a card, a stablecoin address. The agent shouldn’t care. It should say, “Pay this person this amount,” and within seconds, funds arrive—anywhere, anytime.

Agents have their own inboxes, some even have phone numbers—but currently, AI economies lack banks. Agents have no accounts. For agents to have accounts, the underlying funds must:

Who Will Build the Agent-Native Payment Layer?

Payout infrastructure will spawn a new generation of giants, with several reaching market caps of hundreds of billions. From the perspective of those who win the agent race, this is merely a rounding error.

The difference lies in monetization. At the payment layer, margins are 1 to 300 basis points—0.01% to 3%. But an agent deciding what buyers purchase and from whom is far more valuable to sellers than mere capital movement. Companies running these agents will be compensated much like today’s marketplaces and ad platforms.

This doesn’t diminish the importance of payments within agent platforms. On the contrary, their payment teams face some of the toughest challenges in the company: stitching together licenses, banks, card networks, local payment rails, compliance into a secure, usable, globally deployable, buyer-seller-friendly experience. The underlying capital layer will become as reliable and essential as internet protocols.

For those who master the agent-native payment layer, this is a massive opportunity.

This Time, It Might Not Be Too Early

The first wave of agent-driven commerce moved slowly. I can already hear people saying I’m three times too early. I think money is a big reason. Agents have long relied on payment flows built for humans—each purchase ending with a card form and confirmation button.

This time, models are good enough, and laws are written. What’s missing is money that agents can use without screens. Whoever builds it will create the foundation all agents depend on. If I’m once again early—I accept it.

Written by: David Marcus, Former CEO of PayPal, Head of Payments & Crypto at Meta, Translated by: AididiaoJP, Foresight News