Skip to content

Agentic Payments: How AI Agents Pay, and What Has to Be True First

Agentic payments let AI agents pay within limits their owners set. The two ways agents pay today, what each needs first, and who can use them now.

Industry Trend
“Agentic Payments” — hundreds of small glass fragments streaming into one glass coin, how AI agents pay and what has to be true first

Agentic payments are payments an AI agent makes on behalf of a person or business, within limits they set. Today an agent pays in one of two ways: it carries a scoped payment token through a merchant's checkout on card rails, or it pays per request from a wallet after a server answers with HTTP 402, usually in stablecoins.

Neither way works until a few things are in place: the owner's limits, a credential the agent can use safely, a merchant willing to accept agents, a price the rail can carry, and a party who answers for mistakes. The two ways handle each of these differently, and the standards behind both changed owners or versions during 2026.

What "Agentic Payments" Means

On Google's first pages for these searches, the phrase covers three different things: AI agents paying on someone's behalf, AI systems running a payment company's own operations (routing, fraud checks, disputes), and agents getting paid for work they do. This article is about the first.

The term also sits next to agentic commerce. Visa introduced its agent program with the line that it "enables AI to find and buy" (Visa, April 2025). We read agentic commerce as the whole job of finding, choosing and buying, and agentic payments as the payment step inside it.

Two Ways an Agent Pays

Checkouts were built for people. The FIDO Alliance, which writes authentication standards, says today's models "were designed for direct human interaction, not delegated, agent-initiated actions" (FIDO Alliance, April 2026). Two answers have taken shape.

Checkout with a tokenPay per request
Where the agent paysA merchant's checkoutAn API or web resource
What it carriesA network or processor tokenUsually a signed payment from a wallet
Usual moneyCardsStablecoins
ExamplesVisa, Mastercard, Stripe SPTs with ACP, AP2x402, MPP

Way 1: A Scoped Token Through the Checkout

The agent goes through checkout much like a shopper, but it never holds the card number. Visa Intelligent Commerce replaces card details with "tokenized digital credentials" (Visa). Mastercard Agent Pay introduced Agentic Tokens built on Mastercard's existing tokenization (Mastercard, April 2025). Stripe's shared payment tokens (SPTs) let an agent start a payment "using a buyer's permission and preferred payment method, without exposing credentials" (Stripe).

A commerce protocol carries the checkout itself. Stripe developed the Agentic Commerce Protocol (ACP) with OpenAI, and its current docs list Stripe, OpenAI and Meta as the creators (Stripe docs). Under ACP the business stays the merchant of record. Stripe also supports the Universal Commerce Protocol (UCP), which is compatible with Google's Agent Payments Protocol (AP2).

AP2 ties the payment to what the user approved with two signed mandates: a Checkout Mandate for the items and purchase details, and a Payment Mandate for the payment instrument (AP2). In UCP, a payment mandate is scoped to a hash of the checkout, so the token cannot be replayed and the amount cannot be changed (UCP).

Way 2: Pay Per Request From a Wallet

Here the agent calls an API or another web resource directly. If the request arrives without payment, the server answers with HTTP 402 and the agent pays and retries (x402.org). x402 supports EVM-compatible chains, Solana and others, and stablecoins are its primary use case. Usually a facilitator, a server the merchant picks, verifies and executes the payment; the merchant's own server can also do it (x402 on GitHub).

What Has to Be True First

Five conditions decide whether either way works, and each way meets them with different tools.

  1. The owner has set what the agent may do. On cards, Visa lets consumers set spending limits and conditions, and an SPT can be scoped to one business, limited by time or amount, and revoked at any time. In AP2, an open Payment Mandate carries the user's constraints, such as a budget, for purchases the agent makes alone. On the per-request side, x402's upto scheme authorizes up to a maximum per request, and the seller settles the actual usage under that cap. ERC-4337 smart accounts can run their own validation logic, which is where wallet-level spending rules can live.

  2. The agent can pay without holding the real card or key, and only for what was approved. Tokens and mandates do this on cards. In x402, the agent signs each payment, and every payment scheme must not let the facilitator or the server "move funds, other than in accordance with client intentions" (x402 on GitHub).

  3. The merchant can tell a legitimate agent from a bot. Visa warns that bot detection "can mistakenly block legitimate agentic transactions". Its Trusted Agent Protocol, from October 2025, has agents sign their requests with HTTP Message Signatures so merchants can recognize "trusted agents with commerce intent" (Visa, October 2025). Mastercard said in April 2025 that everyone in the chain would be able to recognize transactions made by agents, and under ACP a business can accept or decline "on a per agent, per transaction, or custom logic basis" (ACP). For agents onchain, ERC-8004, a draft Ethereum standard, gives each agent a portable identifier and leaves payments out of scope (EIP-8004).

  4. The price clears the rail's minimum. On Stripe, each payment method has a published minimum, and a per-call price can sit below it. The worked example below puts numbers on this.

  5. Someone answers when it goes wrong. AP2 names accountability as a question today's payment systems cannot answer: if a transaction is fraudulent or wrong, is it "the user, the agent's developer, the merchant, the issuer, the PSP, or the orchestration layer?" Mastercard said in April 2025 that its program would include a process for agent transactions a cardholder does not recognize, and the Machine Payments Protocol (MPP) is designed so that disputes can be "gradually introduced" (MPP). A developer on Hacker News put the practical side in one line: "Regulators want audit trails" (Hacker News, January 2026). Stablecoin payments settle with finality, so we think dispute handling on that side falls to the terms between buyer, agent and merchant, since card chargeback rules do not reach it.

Worked Example: A Price Too Small for the Rail

Say you charge agents $0.002 per call to your API. As of November 2026, Stripe's published minimums are 0.50 USD for card payments through SPTs and 0.01 USDC for stablecoin payments (Stripe docs). By our arithmetic, one card payment has to cover at least 250 calls, and one stablecoin payment at least 5.

Charging per call at that price needs a way to add small amounts up before settling. On EVM chains, x402 has a batch-settlement scheme that "uses escrow and off-chain vouchers so sellers can redeem many small charges onchain in batches" (x402 on GitHub). The figures above use only published minimums; fees come on top, and x402 itself charges no protocol fee but still leaves "nominal payment network fees" (x402.org).

Who Can Use Which Way Today

As of November 2026, the documents set these gates:

  • Visa, for teams building agents on Visa Intelligent Commerce: production access requires becoming "an approved participant within the VIC ecosystem", onboarding to the Trusted Agent Protocol and related Visa services, and registering in the Visa Agent Registry Service so Visa can validate agent signatures (Visa Developer).
  • Stripe SPTs (public preview): available in the US, Canada and select European countries (Stripe docs).
  • Stripe stablecoin payments: available to businesses in all US states except New York, and on request in 30+ other countries; in both cases Stripe reviews the request before the method goes live (Stripe docs).
  • x402: its documentation warns against assuming that the public x402.org facilitator is "the default production path for mainnet EVM routes", so a seller picks a facilitator (x402 on GitHub).

A developer on Hacker News in March 2026 listed what stopped them: "Card issuers won't respond to individual developers", "Stripe requires 3D Secure for off-session payments", and "E-commerce sites block browser automation" (Hacker News).

For a developer or a one-person company (OPC) that wants to sell to agents, the two ways ask for different things. Selling per request with x402 on your own needs a wallet address to receive payment and a facilitator, or your own server doing the verification. Taking card payments from agents through Stripe needs a Stripe account and profile, in the regions where SPTs are open. Building your own agent on the Visa network adds Visa's onboarding. An agent buying from an ordinary web shop still goes through the checkout way.

Where the Standards Stand

As of November 2026:

  • AP2: Google announced it on September 16, 2025, with more than 60 organizations (Google Cloud). On April 28, 2026, Google donated it to the FIDO Alliance and released version 0.2 (Google). Its core supports cards, with push payments and digital currencies on the roadmap; crypto payments run through the A2A x402 extension launched with it. Some guides still describe the 2025 mandates (Intent, Cart and Payment); version 0.2 uses Checkout and Payment mandates.
  • Verifiable Intent: an AP2-compatible standard co-developed with Mastercard and also donated to FIDO, which "creates a tamper-proof log of user-authorized agent actions".
  • x402: Coinbase contributed it to the x402 Foundation, which the Linux Foundation launched on April 2, 2026 (Linux Foundation). Visa said in October 2025 that it is working with Coinbase on interoperability with x402, and it is among the x402 Foundation's initial supporters.
  • MPP: launched by Tempo and Stripe on March 18, 2026.
  • ERC-8004: still a draft, created in August 2025.

Agentic Payments on QDay

QDay Network is a post-quantum, EVM-compatible network built for agent payments. Its payment layer for agents, Signet, is designed around the conditions above, and each part is coming soon: the owner approves a spending session with a budget, permissions and an expiry; the agent pays per request via x402 in stablecoins; and agents register under ERC-8004. Agent wallets with spending limits are also on the way.

Aevum Testnet is live, and no QDay-specific SDK is needed to build on it; mainnet is planned for 2027 Q1. Each payment an agent makes carries a signature, and QDay's signatures follow a migration path: classical ECDSA today, hybrid ECDSA + ML-DSA-65 in progress on Aevum Testnet, and native post-quantum signing on the 2027 roadmap. How a quantum computer would threaten those signatures, and on what timeline, is in What Is Q-Day?

Get the next post in your inbox

Testnet releases, developer updates and research from QDay.

EssentialNeeded for the site to work. Always on.
Always on
AnalyticsGoogle Analytics, to understand how the site is used.
AdvertisingGoogle Ads and other ad tags, to measure and personalise our campaigns.