A flat card charge would cost more than the sub-cent payment it settles. What the Visa-funded research leaves open is who answers when an agent spends money it should not have.
Visa and the blockchain analytics firm Artemis published a joint report on July 16, Agentic Payments from the Ground Up, arguing that the payments software makes to other software — sums well below a dollar — do not belong on card rails at all. The paper works from the technical rules, shared specifications and live onchain figures beneath agents that buy on their own.
It splits agentic commerce in two. In macro commerce, an agent buys for a person at the scale a shopper spends — booking a flight, running a subscription — amounts that fit without strain onto card rails already in place. In micro commerce, one program pays another over and over in tiny sums, for an API call or a sliver of compute, normally far below a dollar. The research Visa commissioned and funded declines to pick one: cards for what an agent buys through a merchant network as it stands today, stablecoins for the small sums software pays software — and it expects many payments to cross both rails inside a single task.
The economics behind that split are measured, not projected. Payments on x402 and on the Machine Payments Protocol each average less than a cent, and a flat card charge on a transfer that small would dwarf the transfer itself — which is why such networks laid down fresh rails instead of reworking what existed. MPP, a Stripe and Tempo creation the card network also helped shape, went live in mid-March 2026 and moved roughly 115,000 transactions worth around $25,000 in total over the weeks immediately after.
The idea itself is old. Web engineers set aside HTTP status code 402 in 1997 and labelled it "Payment Required," on the expectation that people would routinely pay for what they read online; flat card charges left sub-dollar payments uneconomic and the web took the advertising route instead. The report puts the change down to two developments landing together: agents capable enough to generate constant machine-speed demand for paid resources, and newer blockchains that have driven settlement down to fractions of a cent, making transactions from a cent to a dollar workable for the first time.
The commissioned paper finds that on the card side, Visa Intelligent Commerce, the Agent Payments Protocol and the Trusted Agent Protocol are converging with crypto-native protocols such as x402 and MPP rather than staying rivals to them. MPP can already settle onchain and in fiat using payment tokens common to both, and several protocols born on the card side are adding stablecoin support. For network operators, issuers, acquirers and fintech developers, the live question is not whether agentic commerce arrives but where on the range from cards to stablecoins their infrastructure work belongs.
Cheaper rails settle nothing about who answers for a mistake. Commerce as it has been built assumes a human buyer exercising judgment, someone who can be blamed; once an agent does the buying — perhaps because a hostile prompt steered its spending — nobody can say for certain whether liability sits with whoever set the task, the platform running it, the model's builder or the seller. Existing legal and regulatory frameworks were not written for delegation of this kind, and clear precedent may not yet be available. Chargebacks and dispute resolution were assembled for commerce at human pace, where one order can be identified; with agents transacting thousands of times an hour and paying other agents down a chain, nobody has agreed how to reverse a payment that went wrong.