Part of a series on building cz-agents → under the hood

Thirty Years Waiting for HTTP 402

The status code 402 Payment Required has been in the HTTP specification since the nineties, annotated as "reserved for future use". The future never came — instead of native payments, we got forms with card numbers, subscriptions, and advertising. For humans, that works well enough. For software that is supposed to act autonomously, it is a dead end: an agent cannot fill out a captcha, navigate a 3-D Secure challenge in a mobile app, or sign a subscription agreement.

Agentic payment has a straightforwardly describable requirement: a machine needs to pay a small amount, immediately, without a human at the keyboard — but with that human's prior consent and within their defined limits. And it is precisely in that "but" that the entire problem hides. This is not a technology question — it is a question of identity, mandate, and accountability.

x402: Payment as Part of the HTTP Request

In May 2025 Coinbase published an open protocol x402, which finally uses the 402 code for what it was designed for (Coinbase, May 2025). The principle is elegant: the client requests a paid resource, the server returns 402 together with payment terms (amount, token, network), the client attaches a signed stablecoin payment — typically USDC or EURC via the EIP-3009 standard, meaning no gas fees — and repeats the request. No registration, no API keys, no invoice. Payment per individual request, down to fractions of a cent.

In September 2025 Cloudflare and Coinbase announced the formation of the x402 Foundation, so that the protocol would not remain a proprietary technology of a single exchange (Cloudflare, 23 Sep. 2025); in spring 2026 the protocol moved under the neutral governance of the foundation. According to aggregated data from spring 2026 the protocol processed over one hundred million transactions, predominantly on the Base and Solana networks, with an average payment in the range of tens of cents (Eco, 2026) — which is precisely the segment where cards with fixed fees have no chance.

The weaknesses are just as visible as the strengths. Payment in a stablecoin is final — no disputes, no refunds if something goes wrong. A European business also has to figure out how to account for USDC transactions, what the MiCA regulation says about them, and what currency risk they carry — even if small for stablecoins. And above all: x402 solves value transfer, not trust. The fact that an agent paid does not yet mean it was authorised to pay, or that it will receive what it ordered.

Cards: Tokenised Mandate Instead of a Card Number

The card associations responded to agents surprisingly quickly, and both with the same logic: don't give the agent a card, give it a token with a limited mandate.

Mastercard Agent Pay, announced on 29 April 2025, builds on so-called Agentic Tokens — an extension of the existing MDES tokenisation service that today powers contactless mobile payments. The token ties the payment instrument to a specific verified agent, a specific set of merchants, and specific consent rules; the agent never holds the actual card number (Mastercard, 29 Apr. 2025).

Visa Intelligent Commerce followed the next day, 30 April 2025: tokenised "AI-Ready" cards, spending limits set by the consumer, and an API suite for agent developers; among Visa's partners it named Anthropic, OpenAI, Microsoft, and Stripe (Visa, 30 Apr. 2025). In October 2025 Visa added the Trusted Agent Protocol — a framework developed with Cloudflare, built on the HTTP Message Signatures standard, by which an agent cryptographically identifies itself to a merchant. Visa announced it at a time when traffic to US e-commerce sites from AI tools had grown 4,700% year-on-year (Visa, 14 Oct. 2025).

The card approach has what x402 lacks: solved cardholder identity, consumer protection including chargebacks, and decades of merchant trust. The price for that is fees that economically kill micropayments, and the fact that the whole system is still running in pilot mode — it is a mandate layer bolted onto infrastructure designed for humans holding a plastic card.

Banks: Slowest, But with the Regulator on Their Side

The third camp is banks and account-to-account payments. This is where it becomes most apparent that agentic payment "is not just technology". The European PSD2 directive requires strong customer authentication (SCA) every time a payment is initiated — but in an agentic scenario the human authorised the agent, while the transaction was authorised by the agent. Regulation has not yet accounted for this delegation, and legal analyses agree that without explicit mandates with limits, counterparty verification, and an audit trail, it will not pass (Taylor Wessing, February 2026).

Europe has, however, laid foundations elsewhere: the instant payments regulation brought mandatory instant euro transfers and Verification of Payee (checking recipient name against IBAN) for eurozone banks from 9 October 2025; for banks outside the eurozone — including Czech ones — the VoP obligation will not apply until 9 July 2027 (ECB / RedCompass Labs, 2025). Instant settlement and verified recipient identity are exactly the building blocks that agentic payments need — only nobody has yet built an agentic mandate layer on top of them.

That it is possible even within a regulated framework was shown by the first movers: in March 2026 Santander and Mastercard announced the first European payment executed end-to-end by an AI agent within a regulated banking environment — though in a controlled setting, explicitly not as a commercial deployment (Santander, 2 Mar. 2026). In June 2026 Worldline, ING, and Mastercard then completed the first European agentic payment in a live production environment (Mastercard, 2 Jun. 2026).

What an Agentic Payment Actually Needs

When you place these three camps side by side, a common checklist emerges — and it is clear that none of the camps covers it in full:

  1. Agent identity. Who is actually paying? x402 knows only a wallet signature; Visa and Mastercard are building registries of verified agents. Without identity, disputes cannot be resolved.
  2. Mandate and limits. A human must be able to say "this agent may spend at most this much, for these purposes" — and the system must enforce that mechanically, not contractually.
  3. Settlement. A stablecoin settles in seconds and globally; a card settles in days through clearing intermediaries; instant SEPA in seconds, but only within Europe.
  4. Dispute resolution. The finality of an on-chain payment is an advantage for micropayments (low cost), a risk for larger purchases. Chargeback is expensive, but it exists.
  5. Regulation. SCA, AML, MiCA, liability for agent actions. No protocol "codes this away" — which is why banks move slowest and why their path is ultimately probably unavoidable for larger amounts.

My working conclusion: there will not be one winner. x402 will take machine-to-machine micropayments (API, data, compute time), cards will take consumer purchases through agents, and banks will take large and regulated flows. The decisive battle is not being fought over money transfer, but over the mandate and identity layer — and today every camp is building its own fence.

Czech Context: SPAYD as a Pragmatic Interim Step

And what about us? Since November 2012, the Czech Republic has its own open standard SPAYD (QR Payment), which the Czech Banking Association adopted as a unified format for encoding payment details in a QR code (CBA, 14 Nov. 2012). This is not an agentic payment — a human still has to scan the QR code and confirm the payment in the banking app. But it is an interesting intermediate step: the machine prepares the payment, the human authorises it.

This is exactly the principle behind my tool payQR — an open generator of SPAYD and EPC (SEPA) payment QR codes under the MIT licence, which can also be used by an AI agent via MCP. The agent reads an invoice, deterministically assembles the payment QR code, shows it to you — and the final word, a fingerprint in the banking app, remains yours. No delegated authorisation, no mandate, no regulatory grey zones. In the terminology above: the identity is yours, the mandate is "one payment I can see in front of me", settlement is an instant Czech payment, and there is no dispute to resolve because you signed nothing in advance.

This is a deliberately modest answer. But until Europe settles how a machine mandate under PSD2 (and the forthcoming PSD3) should look, "agent proposes, human approves" is the only pattern that works in Czech banking today without exceptions and without waiting. Fully autonomous payments are coming — Santander and ING have already shown it works technically within a regulated framework. The question is not whether, but who will hold the trust layer: a foundation managing an open protocol, the two card associations, or banks with the regulator behind them. We will wait a few more years for the answer — and until then, it pays to understand all three camps.


Sources: Coinbase x402 (May 2025) · Cloudflare x402 Foundation (September 2025) · Mastercard Agent Pay (April 2025) · Visa Intelligent Commerce (April 2025) · Visa Trusted Agent Protocol (October 2025) · Santander × Mastercard (March 2026) · Worldline × ING × Mastercard (June 2026) · ECB on the instant payments regulation · Taylor Wessing on agentic AI in payments (February 2026) · CBA QR Payment standard (November 2012)