In June we asked a question that had no clean answer: when an AI agent spends money, who is accountable for the transaction? At the time the honest response was that the infrastructure for autonomous payments was arriving faster than the framework for assigning responsibility. Six weeks later, the infrastructure question has been substantially settled. The accountability question has not moved.
What Actually Shipped
On 14 July 2026 the Linux Foundation launched the x402 Foundation with around forty founding members, including Visa, Mastercard, Stripe and AWS. That membership list matters more than the technology. It spans card networks, payment processors, cloud infrastructure, stablecoin issuers and blockchain platforms, which means the major participants in the existing payment system have agreed on a common way for software to pay for things without a human in the loop.
The protocol itself is deliberately unremarkable. It revives the long-dormant HTTP 402 status code as a payment handshake embedded in ordinary web requests. An agent requests a paid resource, the server responds with 402 and machine-readable payment terms — the price, the accepted stablecoin, the settlement network and the destination wallet — and the agent pays and retries. There is no new user interface and no new negotiation model. That is the point. It slots into infrastructure that already exists.
Adoption figures should be read carefully, because different sources count differently and the reporting period matters. By April 2026 the Coinbase-led protocol had reportedly processed on the order of 165 million agent transactions and roughly $50 million in cumulative volume across about 69,000 active agents. By early July, reporting around Stripe and Cross River Bank's bank-grade single-use card issuance for agents put cleared x402 transactions at approximately 160 million. The precise number is less instructive than the ratio it implies: an enormous count of very small payments. This is machine-scale commerce — metered API calls, data access, compute — not consumer checkout.
Authorization Is Not Accountability
The emerging stack separates two concerns. AP2 handles authorization and mandates: the rules under which an agent is permitted to act. x402 handles settlement: moving the stablecoin once payment is due. The two are described as complementary, with x402 acting as the settlement extension for the authorization layer above it.
That separation is sound engineering, and it resolves the question that was blocking deployment. An agent can now prove, at the moment of payment, that it holds a valid mandate. The merchant can verify it. The payment settles.
But proving permission at the moment of payment is not the same as being able to answer for the transaction afterwards. Accountability is a question asked later — during a dispute, an audit, a regulatory examination, or a wind-down — and it takes a different form. Not "was this agent allowed to pay?" but "who authorized this specific expenditure, under which policy version, running which model, and can that account be trusted by someone who does not trust the operator?"
The Question an Auditor Will Actually Ask
Consider a firm that discovers, months after the fact, an unexpected pattern of agent spending. Reconstructing what happened requires assembling several records that today live in different places and under different control. The mandate that authorized the agent sits with whoever issued it. The policy in force at that moment sits in a configuration system that has since been updated. The model version that made the decision sits with the model provider. The settlement record sits on-chain. The link between them — this payment came from that decision under this policy — usually sits in application logs owned by the operator being asked to explain itself.
Only one element of that chain is independently verifiable. The stablecoin transfer is on a public ledger. Everything that gives the transfer meaning is in operator-controlled storage, which is exactly the structure we examined in the provenance work: a record whose integrity depends on the honesty of the party with the most to lose from an unfavourable finding.
This is not a hypothetical concern for a fringe use case. Agentic payments are being adopted precisely because they remove human review from routine spending. Removing human review removes the incidental audit trail that human review generates — the approvals, the emails, the sign-offs. The evidentiary record has to be designed in deliberately, because nothing else is producing it.
What the Record Has to Survive
The requirements are the ones this research has arrived at repeatedly from different starting points. The record must be tamper-evident, so that alteration after the fact is detectable rather than merely prohibited. It must be independent of the operator, because its purpose is to be credible to parties who are not obliged to take the operator's word. And it must be final in a way that holds up over the period during which questions can still be asked, which for a regulated firm is measured in years.
None of that requires putting agent decisions themselves on a public ledger, and it should not. What it requires is that the commitments — the mandate, the policy version, the model identifier, the decision that triggered payment — are anchored in a way that lets anyone verify later that the record produced in response to a question is the same record that existed at the time.
Where Mintlayer Fits
Mintlayer is a Bitcoin Layer 2 built for asset issuance and settlement, anchoring to Bitcoin's Proof-of-Work chain rather than replacing its security model. Two properties matter for this problem specifically. Assets and commitments inherit the UTXO structure, so supply and ownership are explicit at the protocol level rather than held in mutable application state that can be rewritten by the same logic that wrote it. And the smart contract model is deliberately non-Turing complete, restricting operations to issuance, transfer and settlement, which keeps the surface area of a record-keeping system small.
Mintlayer Web Services provides the issuance and settlement infrastructure for institutions that need agent activity to leave a record an auditor can trust. The settlement leg of agentic commerce is now a solved problem with the payments industry behind it. The evidentiary leg is not, and the firms deploying agents at scale are accumulating an obligation to explain transactions they currently cannot reconstruct.
The rails arrived first. They usually do. The accounting always follows, and it is generally cheaper to build it before the first question is asked than after.
This article is for informational purposes only and does not constitute investment advice.
Mintlayer Web Services provides Bitcoin-native issuance and settlement infrastructure for institutions. Learn more →
