Two weeks ago a founder in our portfolio showed me an AI agent buying something with no human in the loop. Not a keynote demo rigged to work once. He shared his screen on a Tuesday call, kicked off a task, and the agent paid a data API in a stablecoin, mid-run, in under a second. There was no card number, no checkout page, and he wasn't even watching the payment clear. He was watching the task finish.
That small moment is the thing the entire payments industry has been bracing for, and in June it stopped being a lab trick. On June 7, Forbes ran a piece with a headline that says the quiet part out loud: Visa, Mastercard and Coinbase are fighting over how AI agents pay. Within the same few weeks, Mastercard launched Agent Pay for Machines, Visa unveiled a set of AI and stablecoin tools for agentic commerce, and Coinbase shipped "Coinbase for Agents," letting assistants like ChatGPT and Claude trade and spend from your account.
So now an agent has two rails it can use to move money: the tokenized card and the stablecoin. Yet almost everyone is covering this as a fight between the two. I think that framing is going to age badly. The rail is not the prize. The bridge between the rails is.
What I've Been Seeing
Start with the stablecoin side, because it's the one most people underestimate. The x402 protocol, a way for a machine to pay for any online resource with a stablecoin (no account, no human click), went from a curiosity to real infrastructure fast. By late April 2026 it had crossed roughly 69,000 active agents and 165 million transactions, and across Base, Solana and BNB Chain it now clears somewhere around $600 million in annualized volume. Stripe wired it into its payments API. Google's Agent Payments Protocol references it for agent-to-agent settlement. Cloudflare built a monetization gateway on top of it. None of that needed a card network's permission.
Now look at what the card networks actually did, because it's more interesting than "they're defending the moat." Mastercard agreed in March to buy the stablecoin platform BVNK for up to $1.8 billion. Visa put stablecoins on stage at its own payments forum. And here's the tell I keep pointing to: the x402 Foundation's launch roster includes Visa, Mastercard, Circle, Stripe, Google, AWS and Shopify: 22 organizations sitting inside the same standard.
Read that again. The card networks joined the stablecoin standard's governing body. Nobody with real distribution is betting on a single rail.
The Bridge Thesis Applied
Here's the pattern I keep coming back to, and it's the reason I read this news the way I do. In every major payments shift I've watched over 20 years, value did not accrue to the rail. It accrued to the layer that sat on top of the rails and made the choice invisible. Think about how Stripe actually won. Not by inventing a card network, but by making every card network feel like one clean API that a developer never had to think about twice. Plaid is the same story from a different angle: it never became a bank, it just bridged thousands of them into a single connection anyone could build on. In both cases the rail turned into a commodity, and the connective layer sitting on top of it turned into the real business.
An AI agent makes this even sharper, because an agent doesn't want a rail at all. It wants to finish a task. Whether the data API gets paid in USDC on Solana or on a tokenized Visa credential is not a preference the agent holds. It's a routing decision someone else should make on its behalf, based on cost, speed, and whether the counterparty accepts it. That "someone else" is the whole game. Whoever owns the layer between an agent's intent and final settlement (handling identity, spending limits, authorization, and routing to the right rail) owns the relationship the money flows through.
This is the bridge-builder thesis in about as pure a form as I've seen it. At ChainGPT Labs, where I serve as Investment Director, the abstraction we keep circling back to is exactly this: not "build an agent a wallet," but build the connective tissue that lets an agent transact across whatever rail the moment requires. I wrote a while back in Bridge Notes #2 that AI x Web3 is really three markets, not one. Agentic payments is the clearest example of the market where the two worlds actually have to touch, and the point of contact is worth more than either side.
Where the Analogy Breaks
I want to be honest about where my own framing gets thin, because the Stripe comparison is seductive and it isn't perfect.
The abstraction thesis assumes rails commoditize. They might not — at least not evenly. Visa and Mastercard carry something stablecoin rails haven't replicated yet: a mature apparatus for fraud, disputes, chargebacks and recourse. When a human buys the wrong thing, there's a well-worn path to get the money back. When an autonomous agent misfires at machine speed and drains a budget on a bad instruction, who reverses it? Card networks already have an answer to that question. Stablecoin rails, for the most part, still don't. If regulators decide agent payments must run through KYC'd, reversible channels, the neutral-bridge story weakens and the incumbents keep the toll booth after all.
The other unsolved problem is identity and liability. An agent isn't a person, and "the agent spent it" won't hold up when something goes wrong. Until there's a clean answer to who is accountable when an agent overspends, some of the value I'm attributing to the bridge layer may sit, for a while, wherever the legal recourse lives instead.
What Founders Should Do With This
If you're building in this space, the first thing I'd do is stop pitching me an agent wallet. I've now seen a lot of them, and the wallet is not where the defensibility is. The authorization and policy layer is where it lives: the part that decides what an agent is allowed to spend, on what, up to what limit, and across which rail. That's the piece that's genuinely hard, genuinely sticky, and mostly unbuilt.
Second, pick your altitude on purpose. There are really three places to stand: the rail itself, the abstraction over the rails, or the application on top. Most teams drift between them and end up defensible at none. Decide.
Third, assume multi-rail from day one. Design for x402 and card tokens together, because the enterprise buyers I talk to are not going to standardize on one for years, if ever. And treat the recourse gap (refunds, disputes, an agent's version of a chargeback) as a product to build, not a footnote to wave away. Frankly, the team that makes agent payments reversible in a way businesses trust may end up mattering more than the team that made them fast.
In payments, the rail is never the moat. What matters is whoever the money has to pass through to reach it — and in agentic commerce, that gatekeeper is the bridge between the two rails, not either rail on its own.
Questions I Keep Getting
What is x402? It's an open protocol that lets software, including an AI agent, pay for an online resource with a stablecoin automatically, without a stored account or a human clicking "buy." It reuses the old HTTP 402 "Payment Required" status, and by mid-2026 it was clearing hundreds of millions of transactions across Base, Solana and BNB Chain.
Will AI agents pay in stablecoins or on cards? Both, and that's the point. The serious players (Visa, Mastercard, Coinbase, Stripe, Google) are hedging across both rails simultaneously. Betting on a single winner is the wrong bet. The layer that lets an agent use either one without caring is the position I'd want to hold.
Who wins the agentic payments war? Probably not "a rail." If history rhymes, the durable value lands with whoever owns the authorization, identity and routing layer between the agent and settlement, the same spot Stripe took over cards. That's a bridge, not a rail, and it's the seat I'd tell a founder to fight for.