PRIM3 Brief #17: Agentic Payments Are Real — the Rail Is Commoditized, the Mandate Layer Isn't
Between May 2025 and April 2026, autonomous AI agents settled about $73 million across 176 million transactions. Almost none of it touched a card. That's the headline from Keyrock's agentic-payments report, produced with Coinbase, Tempo, and Virtuals: stablecoins have quietly become the default settlement layer for machine-to-machine commerce, with USDC handling 98.6% of the flow (Cointelegraph).
Look at the shape of that volume and you see why cards were never in the running. The median transaction ran between one and ten cents, and roughly 76% of it fell below $0.30, the floor where card interchange fees stop making economic sense (crypto.news). An agent paying a fraction of a cent for an API call can't be routed through a network built for $40 purchases. On a Layer 2, that same settlement costs about $0.0001. So the rail question, for the micropayment leg of agentic commerce, is effectively answered.
What almost no one is pricing correctly is where the value accrues once it is. The consensus attention is aimed at agent tokens and agentic-commerce front-ends. PRIM3's read is narrower, and we think more useful: the settlement rail is commoditizing in real time, and the defensible, crypto-native layer is the one sitting above it — the mandate and identity layer that decides which agent is allowed to spend what, on whose authority. That's where we're looking for the next check.
The State of Agentic Payments in Mid-2026
Agentic commerce stopped being a slide in a keynote and started being infrastructure this year. Stablecoins processed roughly $46 trillion in 2025, a figure that put them past Visa's annual volume (CoinGecko), and the agent economy is now grafting onto those rails. The 176 million agent transactions in Keyrock's dataset ran over x402, the open payment standard Coinbase built with the Ethereum Foundation for on-chain settlement.
The numbers people keep quoting are the forward ones. Gartner projects AI agents could intermediate $15 trillion in purchases by 2028; McKinsey pegs retail agentic commerce at $3–5 trillion by 2030 (Spark). Round those down by an order of magnitude and it's still one of the largest new payment flows in a decade. And that is exactly why every incumbent is now in the room. Coinbase, Stripe, and Visa have each moved to become the payment layer for agents (CoinDesk).
The moment Stripe and Visa race to own a sector's rail, that rail stops being the opportunity.
It becomes the loss leader.
Why the Consensus Read Is Incomplete
Two things about the current data should reset where founders and investors are looking. First, USDC handles 98.6% of agent settlement. That concentration is so extreme the rail is already a near-monopoly commodity. Building the 51st stablecoin, or a marginally cheaper settlement chain, is competing on the one dimension the market has already resolved. We flagged the systemic side of that concentration in Brief #10; the commercial implication here is simpler. You don't get paid for being a slightly better version of the thing already winning 98.6% of the volume.
Second, the settlement standard is open. x402 was designed with the Ethereum Foundation and is being adopted across the industry, and open standards are precisely the layer where margin goes to zero by design. TCP/IP made a lot of people rich. None of them sold TCP/IP.
The agent-token trade has the same problem from the other end. Most "AI agent" tokens are a narrative wrapped around a rail that a foundation and an exchange are standardizing for free. The pattern rhymes with what we wrote in Brief #4 about AI capex spilling into crypto: the spend is real, but it accrues to a specific sub-layer, not to whichever token is loudest about the theme.
Who's Liable When an Agent Overspends?
Here's the question the volume charts skip. An autonomous agent holding a wallet and a spending budget is a new kind of economic actor, and the moment one gets prompt-injected into a bad purchase, or simply misreads a task and drains a limit, someone has to adjudicate it. Cards spent forty years building chargebacks, identity checks, and liability rules for exactly this. Agentic payments have almost none of it yet.
That gap is the actual product surface, and the industry's own architecture already points at it. Google's Agent Payments Protocol (AP2), launched in September 2025 with 60+ partners including Mastercard, PayPal, Coinbase, and American Express, doesn't define a rail at all. It defines authorization: three signed "Mandates" (Intent, Cart, and Payment) carried as W3C Verifiable Credentials that prove a user allowed a specific agent to spend on their behalf (Google Cloud). Visa's Trusted Agent Protocol does the identity half, verifying that the agent is who it claims to be. x402 sits underneath all of it as the settlement extension (TechTimes).
Read that stack top to bottom and the rail is one line at the bottom. Everything valuable — the mandate, the credential, the identity, the revocation, the audit trail — lives above it. And most of those primitives are things crypto is unusually good at: verifiable credentials, programmable spending limits, on-chain identity and reputation, revocable authority. This is a crypto-native problem that the card networks are solving with card-network tools, which leaves an opening for teams that solve it natively.
What This Means for Founders
If you're building in agentic payments, the single most useful reframe is this: stop trying to move the money and start trying to govern it. The money already moves, for $0.0001, on a standard nobody owns.
The action items follow from where value actually sits.
- Don't build a rail. If your only moat is being cheaper than USDC-on-an-L2 at a hundredth of a cent, you have no moat. The rail is a commodity with a 98.6% incumbent. Assume it and build on top.
- Build the mandate layer. Verifiable authorization means provable, revocable, limit-bound credentials that let a user delegate spending to an agent. It's under-built, and it's standards-aligned with AP2. This is the layer with pricing power, because it's the layer that carries liability.
- Treat agent identity as first-class. An agent needs a durable, verifiable identity and a reputation that travels across services, the way Moca Network (a PRIM3 portfolio project building privacy-preserving identity across chains) and Verida (our portfolio personal-data layer for AI agents) are already building for the human and data side. Whoever owns the agent's passport owns a chokepoint.
- Instrument the dispute path. Micropayments at 176-million scale will produce errors at scale. The team that makes agent overspend legible, attestable, disputable, and insurable is selling the thing the volume charts assume already exists.
The quantitative filter we give founders in this space: if your deck's core metric is settlement cost or throughput, you're pitching the commodity layer. If it's authorization coverage, dispute resolution, or identity reach across agent frameworks, you're pitching the layer that compounds.
Where PRIM3 Is Placing Bets
PRIM3 invests at the intersection of AI, DePIN, and Web3, and agentic payments is close to the purest expression of that intersection we've seen this cycle. When we evaluate a team here, we weight three things heaviest: whether they're building above the rail rather than on it, whether they treat the agent as an economic actor that needs identity and controls rather than just a wallet, and whether their design is standards-aligned with where AP2 and x402 are actually converging. Infrastructure-adjacent bets like Oobit (crypto payments at Visa-accepting merchants) and Kima Network (settlement without bridges) sit near this flow. But the checks we're most interested in writing are one layer up, in authorization, identity, and the liability plumbing that agentic commerce hasn't built yet.
The rail was the easy part, and it's done. The mandate layer is where the next durable companies in agent commerce get built. That's where PRIM3 is spending its attention this quarter. More on our thesis and portfolio at prim3.vc.
FAQ
What are agentic payments? Agentic payments are transactions AI agents execute autonomously on a user's behalf, paying for compute, data, APIs, or goods without a human at the checkout. In 2026, stablecoins became the default settlement rail: agents settled about $73 million across 176 million transactions between May 2025 and April 2026, 98.6% of it in USDC, per Keyrock. The median transaction was one to ten cents.
Why are stablecoins the default rail for AI agent payments? Because the transactions are tiny. Around 76% fall below the $0.30 floor where card fees stop working, and an agent paying fractions of a cent can't route through a card network. On a Layer 2, that settlement costs about $0.0001, which is why crypto rails, not cards, carry the micropayment leg of agentic commerce.
What is the investable layer in agentic payments? Not the rail. With USDC at 98.6% of volume and open standards like x402 and AP2 defining how agents pay, settlement is commoditizing and won't capture margin. The defensible, crypto-native layer is authorization and identity: the mandates proving a user allowed an agent to spend, verifiable agent identity, programmable limits, and the liability trail when an agent overspends.
That's the layer PRIM3 is underwriting this quarter. More on our thesis and portfolio at prim3.vc.