1. What is x402, and why does it lack dispute rights?
In a column published on October 28, 2026, Albert Castellana, co-founder and chief executive of GenLayer Labs, argues that x402, the standard AI agents use to pay for things online, has no way to dispute a charge. Visa, Mastercard and American Express now sit in its top governance tier, even though a settled stablecoin payment cannot be reversed.
Coinbase published the x402 protocol in May 2025. It lets software pay for web resources directly, using stablecoins settled onchain. Settlement takes hundreds of milliseconds and costs a fraction of a cent, which makes tiny machine-to-machine payments practical. Castellana says the decision to leave out dispute rights was deliberate and, in his view, reasonable at a time when usage was sparse.
His concern is what happens as volumes grow. Once a payment settles onchain, there is no issuer to call and no card network to request a chargeback. The column asks what recourse exists when the money moves correctly but the buyer receives something wrong.
For readers new to the topic, an AI agent is software that can act on a user's behalf, including by buying data, computing time or other online services without a person approving each purchase. A standard such as x402 gives those agents a common way to pay, which is why its governance and its safety features matter to a wide set of companies.
The column frames the issue as one of trust rather than technology. Payment networks succeed when users believe that a failed or disputed transaction has a remedy, and Castellana argues that the same expectation will eventually apply to software buyers and the companies that deploy them.
2. Who governs the standard?
In July, the Linux Foundation launched operations of the x402 Foundation, the body that governs the standard. It has 17 premier members, the top governance tier, and Visa, Mastercard and American Express are among them. Castellana points out the irony that all three companies help govern a system with no formal dispute process.
Mastercard said at the launch that internet-native payments need the same secure, trusted and interoperable foundation as today's digital commerce. Castellana notes that trust in card payments has rested on a specific mechanism for more than 50 years, and that the same mechanism is missing from x402.
Governance tiers in standards bodies generally determine who helps set the rules. A seat in the top tier gives a company influence over how the standard evolves, which is why Castellana highlights the card networks' membership. They have long operated dispute systems, so their presence invites the question of whether any of that experience will be built into x402.
3. Why do card payments have dispute rights?
United States law has protected the right to dispute a credit card charge since 1974, when the Fair Credit Billing Act was passed. The protection helped make consumers willing to buy from strangers online: if an item never arrives, the buyer can contact the bank and recover the money.
Castellana observes that these protections assume a person who notices a problem and complains. In x402, the buyer is often software. An AI agent does not read a monthly statement or notice a missing delivery, so the safeguard that card networks built around human attention has no obvious equivalent in an agent-driven payment flow.
The 1974 law is part of the reason an online purchase from an unknown merchant is a manageable risk for a consumer. The right to dispute shifts part of the risk from the buyer to the card issuer and merchant, and it creates an incentive for sellers to deliver what they promised. Without it, the buyer carries the full risk of a bad transaction, a burden that is difficult to accept at the scale of autonomous software.
4. What did researchers find in live x402 deployments?
A July study by researchers at EPFL and Zhejiang University was the first systematic security study of live x402 deployments. The team tested 15 facilitators, the intermediaries that verify payments and push them onchain. Every facilitator broke at least one of eight safety rules the researchers had defined.
The study also measured failures at the settlement stage. On Base, close to 2% of settlement attempts reverted in late 2025, which Castellana puts at roughly 1.9 million payments. A reverted settlement means the payment did not complete, so a seller that had already acted on the payment could be left unpaid.
That risk is visible in Coinbase's own seller tools. All seven of its official reference kits for sellers lack a way to reverse actions the server has already taken when settlement fails. The Flask kit, through version 0.2.1, releases goods once the payment proof verifies, before settlement is confirmed. If settlement then fails, the seller has delivered the goods and received nothing.
A facilitator is the intermediary that checks a payment before a seller releases goods and then submits the transaction to the blockchain. Because it stands between buyer and seller, a flaw there can cause the two sides to disagree about whether a payment actually happened. The finding that every one of the 15 facilitators tested broke at least one safety rule shows that the weakness is not limited to a single provider.
Taken together, the study's results and the reference-kit behavior describe a stack in which several layers can fail independently. A buyer's agent, a facilitator, the blockchain and the seller's server each play a part, and the order in which they act determines who bears the loss when something breaks.
5. What if the payment settles but the agent gets the wrong thing?
Castellana treats the settlement bugs as the lesser issue, since early infrastructure often has them. The deeper problem, he writes, is that payment, delivery and settlement can disagree. The harder case is a payment that settles correctly while the agent receives something wrong, incomplete or useless.
He stresses that this is not an argument against x402. Payment finality is not commercial finality: a blockchain can show that money moved, but it cannot determine whether the data was correct, whether a service was delivered or whether either party met its obligations.
The standard does support card rails alongside stablecoins, and card payments keep their protections. Yet 97% of observed x402 transactions run on Base and Solana, where stablecoin payments are final. In practice, most agent payments today fall outside the card-style safeguards.
The distinction between payment finality and commercial finality is the core of the column. A transaction on a blockchain is final in the technical sense: it cannot be undone. But a purchase involves a promise about what the buyer will receive, and nothing in the payment record verifies that promise. When the buyer is an agent, the problem is amplified, because it may not recognize that anything went wrong.
Castellana gives no estimate of how often an agent receives a wrong or useless result, and the column does not claim that such failures are common today. His point is structural: the system has no designated place for such a complaint to go, so any loss falls on whoever happens to hold the risk at the moment of settlement.
6. Who would absorb the losses?
The column notes that card networks spent five decades answering the dispute question, and each answer cost someone money. Several questions remain open for agent payments: who absorbs the losses, who decides whether a service arrived, how long a buyer has to complain, and what counts as evidence.
Scale complicates the answers. At a tenth of a cent per transaction, disputes cannot involve lawyers, and most cannot involve humans at all. A system built for micropayments by software needs a resolution process that operates at the same speed and cost, which is the gap Castellana wants the industry to address.
The choice facing the industry is therefore about design rather than principle. Fast, cheap, irreversible payments are what make x402 useful, while recourse is what makes a payment system trustworthy to cautious users. The question Castellana raises is how to combine the two, for example through automated review that matches the speed of the payments themselves, without giving up what makes micropayments workable.
The question of evidence is especially hard in an automated setting. A human buyer can describe what went wrong, attach a screenshot or explain an expectation. An agent working through an application interface may leave only a record of the request and the response, and judging whether that response was adequate requires a standard that both parties accepted in advance.
The same problem affects timing. Card disputes typically allow a window in which a buyer can raise a complaint, and the length of that window reflects a balance between protecting buyers and giving sellers certainty that a sale is complete. An agent payment standard would have to choose its own balance, and the choice would determine how much risk sellers and buyers each carry.
7. What happens next?
Castellana says agent payment volumes remain small enough that a bad month would be survivable. The companies in the top governance tier, he adds, have deep institutional knowledge of what dispute processes cost and what they provide. The open question is whether they will build a machine-native replacement before losses force the issue.
His closing line sums up the argument: the industry has spent years making money programmable, and it now needs to make recourse programmable too. The column carries the author's views, not those of any publisher or affiliate.
The column's timing reflects the early stage of the market. Because volumes are small, the cost of any single failure is limited, and Castellana treats the present as a window in which design choices can be made before large losses force them. His implicit request is for the governance members to use their experience while the stakes are still modest.