*Published July 8, 2026* --- *AI agents won't settle on cards or on stablecoins. They'll use both. The rails are segmenting, not converging, so the layer that lasts is the trust and identity one riding on top of whichever rail clears a given transaction.* Two stories get told about building infrastructure for AI agents. Both of them mislead. The first is the gold-rush pitch. Agents will need everything humans needed, the argument goes, so there are twenty-odd open categories sitting there waiting for a founder: identity, reputation, escrow, dispute resolution, liability, work verification. Pick one and build. The second is the one I fell for, the correction to the first: the incumbents already shipped this, so the ground is taken. By mid-2026 Visa has a Trusted Agent Protocol, Mastercard has Agent Pay, Google has an agent payments protocol, and every one of them authenticates an agent against a real cardholder, inheriting the chargeback and dispute machinery that took decades to build. If the expensive layers are already occupied, the honest thing to tell a builder is to go home. When I pushed on that second story, it broke. Not because the incumbents aren't real, but because "occupied" was the wrong word. ## The war everyone is watching The visible part is a fight about rails. Visa and Mastercard are betting agents will pay the way you already do, over tokenized card credentials, with all the chargeback and dispute plumbing that already exists. Coinbase and a16z are betting against exactly that: agents are a different kind of buyer, they need rails built for them, and stablecoins are those rails. a16z's version runs like this: today's agents are tourists in a bazaar, overpaying at retail with cards, when what they are is businesses that should be moving on wholesale terms. Read the coverage, Forbes included, and it comes across as a straight two-sided fight over how AI agents will pay. So the builder's question looks like it writes itself: work out which rail wins, then build on that one. I think it is the wrong question, because it assumes there is going to be one. ## The rails are segmenting, not converging Look at where the transactions are landing, though, and the fight starts to look more like a division of labor. Card networks are taking the consumer-facing side of agent commerce, the part where chargebacks and consumer protection matter and a human is on the hook when something goes wrong. Stablecoins are taking the machine-to-machine side, where sub-cent fees and programmability are what decide things and there is no merchant on either end to protect. That is the split Mastercard's own policy team, American Banker, and Forbes had all landed on by the middle of 2026. The rails aren't converging on one winner; they're dividing the labor between them. The clearest tell is what the incumbents do rather than what they say. Mastercard's Agent Pay program lists Coinbase, Stripe, Solana, and Polygon as partners. A company that expected to win outright would not be building on-ramps for its supposed challenger. None of this would be new. Credit, debit, and ACH never resolved into a single winner either; they split by use case and have coexisted for decades. Betting on one agent rail to wipe out the others misreads how these fights tend to end. ## The layer worth owning If no rail wins outright, the layer worth owning is the one that doesn't care which rail does. That's trust and identity: proving an agent acts for a real principal, that it has authority to spend, that it did the job, that it can be trusted next time. That holds whether the money moves over a card or a stablecoin. The obvious objection is that the incumbents just take this layer too. They can't, and the reason is structural. Visa can't credibly vouch for a job that settled in stablecoins; Coinbase can't underwrite a card dispute. A trust layer everyone leans on has to work the same no matter which rail clears the payment, and only something that sits above both rails can do that. That neutrality is the moat, and neither incumbent can copy it without abandoning the rail it is built on. The money is already moving there, quietly, while the noise stays on the rails. Skyfire, which raised roughly nine million dollars, doesn't call itself a payments company; it calls itself "the agent trust stack." Catena Labs raised a thirty-million-dollar Series A, co-led by a16z's crypto arm, to build banking and governance for agents. Both partner across rails instead of picking one. There's precedent for the bet. When Visa and PayPal were built, the franchises that lasted captured identity and trust while banks and processors moved the actual money. ## The honest caveats This is a mid-2026 read, not a settled result, and two things could change it. Segmentation is still a direction, not a destination. It is where the evidence points right now, and it could tip if one rail turns out to serve both segments well enough. And the demand isn't here yet. A Coinbase-backed agent-payments protocol admitted as much this spring: the plumbing is being laid ahead of the traffic. The open question underneath the whole thesis is whether the trust layer stays an independent business or gets absorbed by whichever rail ends up strongest. Skyfire and Catena partner with Mastercard today. Partners get acquired. So I'm not predicting the trust layer wins. The claim is smaller, and I think more useful: the real contest was never "which rail." It's "what survives whichever rail wins." Almost nobody is asking that one out loud. ## Sources - Forbes — "Visa, Mastercard and Coinbase Are Fighting Over How AI Agents Pay", 2026-06-07 — https://www.forbes.com/sites/digital-assets/2026/06/07/visa-mastercard-and-coinbase-are-fighting-over-how-ai-agents-pay/ - a16z crypto — "Tourists in the bazaar: why agents will need B2B payments" — https://a16zcrypto.com/posts/article/against-b2b-payments-stablecoins/ - Mastercard — "Stablecoins, AI and real-time payments are converging" (policy) — https://www.mastercard.com/global/en/news-and-trends/stories/2026/converging-payments-ecosystem-policy.html - Mastercard — "Mastercard launches Agent Pay for Machines" (partners list) — https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html - American Banker — "AI agents will require a brand-new set of payments rails" — https://www.americanbanker.com/payments/opinion/ai-agents-will-require-a-brand-new-set-of-payments-rails - CoinDesk — "Coinbase-backed AI payments protocol wants to fix micropayment but demand is just not there yet", 2026-03-11 — https://www.coindesk.com/markets/2026/03/11/coinbase-backed-ai-payments-protocol-wants-to-fix-micropayment-but-demand-is-just-not-there-yet - Google Cloud — "Announcing Agent Payments Protocol (AP2)" — https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol - Finextra — "Skyfire raises $8.5m to bring autonomous payments to AI agents" — https://www.finextra.com/newsarticle/44621/skyfire-raises-85m-to-bring-autonomous-payments-to-ai-agents - Catena Labs — company profile (Series A coverage) — https://www.linkedin.com/company/catenalabs