Issuing bank cards to machines: A complete record of the twelve-month surge in AI agent payments

On May 7th, AWS installed a stablecoin wallet on its AI Agents running on its cloud. The partners are Coinbase and Stripe. Agents can autonomously discover APIs, negotiate prices, and pay with USDC during task execution. Warner Bros. is already testing it for advanced access to purchase sports events and entertainment content. In the same week, Google and the Solana Foundation jointly released Pay.sh, an open-source gateway that allows Agents to use Google Cloud services on demand with USDC. MoonPay issued a virtual Mastercard debit card, allowing Agents to spend directly from their on-chain wallets at merchants worldwide. Stripe announced its Agent-specific virtual debit card issuance service at its annual Sessions conference. Two days later at Consensus Miami, nearly 1,000 developers flocked to the EasyA hackathon to build agent payment applications, including engineers from Microsoft and Google. The fact that these events occurred in the same week is no coincidence. They represent the culmination of a series of infrastructure developments over the past twelve months. This article attempts to clarify this unfolding story—from technical protocols to business competition, from on-chain micropayments to offline card payments, from proven successful scenarios to unresolved issues—hoping to provide clues for entrepreneurs, investors, and anyone wanting to understand "where AI agent payments stand." An agent without a wallet can do everything except complete transactions. Before delving into any technical details, let's clarify one thing: why payment is the key bottleneck hindering the agent economy. AI agents today can search, analyze, write code, manage calendars, and book meeting rooms. But there's one thing they can't do: spend money. It's not a lack of ability, but a lack of infrastructure. Traditional payment systems—bank accounts, ID cards, credit cards—are all designed for humans. Agents don't have ID cards, can't pass KYC, and can't fill in credit card numbers on checkout pages. When an agent is simply helping you look up information or report the weather, this doesn't matter. But once it can independently complete multi-step tasks like "research → price comparison → order placement → delivery," payment becomes the most critical bottleneck. An agent without a wallet is like a shopper standing at the checkout counter without a bank card—it can do everything, but it can't complete the transaction. This is not a small problem. It's a prerequisite for AI to evolve from an "auxiliary tool" into an "independent economic agent." Coinbase co-founder Brian Armstrong predicted, "Soon, the number of AI agents doing transactions will surpass that of humans. They can't open bank accounts, but they can have a crypto wallet." Visa's Chief Product and Strategy Officer, Jack Forestell, put it more bluntly: "In my two decades in the payments industry, I haven't seen a bigger growth opportunity than Agentic Commerce." If these predictions hold true, who will build this financial pipeline for agents will be one of the most important infrastructures of the next decade. The Five-Layer Spectrum of Agent Payments: From $33 Trillion to $28,000 Daily. The term "Agent payments" refers to vastly different things in different contexts. Before discussing the technology and applications, let's establish a coordinate system. Layer 1: Stablecoins – The Settlement Foundation of the Agent Economy In 2025, global stablecoin trading volume reached $33 trillion, a year-on-year increase of 72%, with total supply exceeding $300 billion. USDC alone processed $18.3 trillion. This is the underlying clearing pipeline for all agent payments – every transaction in the agent economy ultimately goes through this pipeline for settlement. Layer 2: AI-Assisted Shopping – People Pay, AI Helps Choose eMarketer estimates that AI platforms will contribute $20.9 billion to retail spending in 2026. Salesforce data shows that during Black Friday 2025, one in six online purchases was completed with AI assistance. These figures need to be understood precisely: these are purchases made by humans within AI interfaces such as ChatGPT, Google AI Mode, and Microsoft Copilot—AI participates in discovery and price comparison, but the final "purchase" is still made by a human, using the traditional credit card payment method. "AI-assisted human payment" and "agent-driven payment" are two different things. A landmark counterexample: OpenAI's ChatGPT Instant Checkout was launched with great fanfare in September 2025, only to be quietly discontinued five months later. Only about 12 out of millions of Shopify merchants integrated it. Consumers are happy to have AI recommend products, but not willing to pay directly through the chat interface. Layer Three: DeFi Transaction Agents – The Agents Currently Truly “Spending Money” on a Large Scale This is currently the largest agent-driven payment scenario, estimated at billions of dollars, but most people are still unaware of it. These agent payments bypass x402 or MPP protocols, interacting directly with smart contracts through on-chain wallets. During its 14-week public beta test from October 2025 to January 2026, MoonPay Agents saw over 1,000 participants create more than 9,500 agents, executing 187,000 autonomous cryptocurrency transactions. On the Solana network, a single AI agent manages more daily trading volume than the bottom 20% of human retail investors. Kraken, Binance, OKX, and Gemini released a flurry of production-grade Agent trading toolkits between late 2025 and early 2026. OKX's Agent Trade Kit covers over 60 chains and 500+ DEXs, processing 1.2 billion API calls daily. BingX's AI Arena enabled GPT, Claude, Gemini, Grok, and DeepSeek to trade against each other in real-time with a virtual deposit of $10,000. These are not lab demonstrations, but production-grade infrastructure from the world's largest exchanges. The fourth layer: Card organization Agent payments – an extension of Visa and Mastercard. Traditional payment giants didn't build new systems from scratch; instead, they extended their existing credit card networks to Agents. Mastercard's CEO confirmed the "first Agentic transaction" on the network in October 2025. In March 2026, Santander and Mastercard completed the first agent payment in a regulated banking environment in Europe. In late April, Mastercard and Rabobank completed the first in the Netherlands—an agent booked a coffee tasting experience via Mastercard Agent Pay after sending a notification in ChatGPT, without accessing the cardholder's data. Agent Pay officially launched in ASEAN countries in May. Visa has processed "hundreds" of agent transactions, and the Agentic Ready project has expanded to Latin America and Asia Pacific. The transaction volume is still small—Mastercard's "first transaction," Visa's "hundreds"—but they run on a global network that processes over 300 billion transactions annually. Layer 5: x402 / MPP — Payment Protocols Natively Designed for Agents Finally, we come to the x402 and MPP protocols — the core theme of this article, but currently the smallest layer. As of the end of April, the x402 protocol had 69,000 active agents, 165 million cumulative transactions, and approximately $50 million in cumulative transaction volume. On-chain data shows an average daily actual transaction volume of approximately $28,000, averaging $0.20 per transaction, with about half being testing and volume manipulation. All agent payment activity accounts for 0.0001% of the total stablecoin transaction volume. 50 million cumulative transactions vs. 28,000 daily transactions – the tension between these figures is key to understanding the current stage. The infrastructure is in place, but real business demand hasn't caught up. Why are the world's largest technology and financial companies betting heavily on a market with an average daily transaction volume of $28,000? Because they're not betting on today's scale, but on the end result this pipeline leads to. McKinsey estimates that by 2030, agent-driven commerce will generate three to five trillion dollars in global revenue; Morgan Stanley estimates that 10-20% of US e-commerce will be driven by agents, corresponding to a scale of $190 billion to $385 billion; and Gartner predicts that the B2B agent market will reach $15 trillion in transactions by 2028. The huge discrepancies between these predictions are not a problem of analytical quality, but a problem of definition—"AI-assisted," "AI-initiated," and "AI-autonomous" are three completely different categories. For investors, distinguishing between these three categories is the first step in making a judgment. The story of x402 begins with the most fundamental design of the internet. In 1991, the designers of the HTTP protocol reserved a status code: 402, meaning Payment Required. Imagine a day when network resources could be charged directly at the protocol layer—you request a webpage, the server says "pay first," payment is made, and the content is delivered. This day has been awaited for a long time, but it has never arrived because there is neither a sufficiently cheap settlement channel nor a scenario requiring automatic payment. 34 years later, two conditions were met simultaneously: stablecoins provided a virtually free settlement channel, and AI agents provided users who truly needed to make payments 24/7. In May 2025, Coinbase and Cloudflare released the x402 protocol. The logic can be summarized in one sentence: the agent requests resources, the server replies "pay $0.001", the agent signs the payment with USDC, and the resource is delivered. This is completed within one HTTP request cycle. There are no accounts, no subscriptions, and no API keys. Gas fees on the Base network are approximately $0.001 per transaction. x402 deliberately made a design choice: completely separating payment and identity. The protocol only cares about "whether the money has arrived," not "who paid." Cryptographic verification handles everything. This makes x402 extremely agent-friendly—no registration required, no proof of identity needed, just pay and you're good to go. The cost is equally clear: x402 doesn't prevent agents from overspending or making malicious purchases. Security and identity issues are left to other layers to handle. Looking back at the timeline, speed is the most noteworthy variable. In December 2025, x402 was upgraded to V2, adding reusable sessions, multi-chain support, and automatic service discovery—evolving from "the agent paying once" to "the agent paying more than ten times in a complex task." On March 2, 2026, Santander and Mastercard completed the first AI Agent payment in a regulated banking environment in Europe. Card organizations announced their entry with actual transactions. On March 18, 2026, Stripe made its move. Tempo, a payment-specific L1 blockchain jointly invested in by Stripe and Paradigm, launched the Machine Payments Protocol (MPP). MPP shares the underlying HTTP 402 signature mechanism with x402, but additionally supports subscriptions, streaming billing, cancellations, and refunds, and—crucially—it integrates with Visa credit cards. While x402 addresses "agents pay once," MPP addresses "agents continuously pay for a service." On the day of launch, over 100 services integrated the protocol, including Anthropic, OpenAI, and Shopify. On the same day, Skyfire and F5 announced a collaboration to differentiate between legitimate agents and malicious automated traffic at the network edge. F5 handles approximately 60% of global network traffic. On April 2, 2026, Coinbase donated x402 to the Linux Foundation, establishing the x402 Foundation with 22 founding members: governance from Coinbase, Cloudflare, and Stripe, as well as AWS, Google, Microsoft, Visa, Mastercard, American Express, Shopify, the Solana Foundation, Circle, Ant Financial, Adyen, Fiserv, KakaoPay, and Polygon Labs. The protocol is licensed under the Apache 2.0 license, with zero protocol fees and zero vendor lock-in. The significance of this list is self-evident—it encompasses the three giants of cloud computing, two major card organizations, the world's largest e-commerce SaaS, and major stablecoins and blockchain infrastructure. Linux Foundation CEO Jim Zemlin's vision is clear: x402 aims to become the TCP/IP of the payments industry. From an open-source project to being taken over by the Linux Foundation, eleven months. This speed is unprecedented in technology history. Giants Enter the Fray: Each with Their Own Intentions The protocol is the foundation. But what truly makes this irreversible is that large companies have begun to embed it into their core products. Their entry strategies differ, implying different assessments of the Agent economy. AWS chose the most direct approach: installing wallets for Agents on its own cloud. AgentCore Payments allows developers to connect to Coinbase or Stripe Privy wallets, set session spending limits, and Agents can then autonomously pay within the task flow. The first phase focuses on API micro-payments, with hotel bookings and travel procurement already outlined in the roadmap. AWS is betting on its platform entry point—if the agent economy takes off, with a large number of agents running on AWS, every transaction will pass through Amazon's pipeline. Google's entry strategy is even more intriguing. It didn't choose Ethereum or Base, but instead chose Solana as its settlement layer. Pay.sh allows agents to use Google Cloud services like Gemini, BigQuery, and Vertex AI on a pay-per-use basis using USDC. There are no accounts, no subscriptions; payment is the credential. Google is betting on its ecosystem—turning its own AI services into commodities that agents can directly access and pay for. Solana's positioning in this game is very clever. Instead of passively waiting for developers to build it, it actively participates in the x402 Foundation governance, while making Pay.sh compatible with both the x402 and MPP protocols, acting as a compatibility layer between the two competing standards. Solana already handles nearly 65% of x402 transaction volume—it doesn't take sides, it acts as a pipeline. Stripe's actions are the most intensive and noteworthy. It released MPP, participated in the x402 Foundation governance, acquired the stablecoin infrastructure company Bridge for $1.1 billion, provides wallet services for AWS AgentCore through Privy, and launched 288 new products and features at Sessions 2026. The CEO announced: "Stripe is building economic infrastructure for AI." Stripe isn't taking sides, but rather making itself an indispensable pipeline for everyone. Visa and Mastercard take completely different paths. They don't issue new protocols or touch stablecoins; instead, they extend their existing credit card trust networks directly to agents. Visa's Trusted Agent Protocol uses cryptographic signatures to issue credentials to agents, and its Agentic Ready project has expanded to Latin America and the Asia-Pacific region. Mastercard's Agent Pay has launched on three continents, and its Verifiable Intent standard, developed in collaboration with Google, packages consumer identity, instructions, and transaction results into tamper-proof records. American Express launched Agent Purchase Protection—the industry's first commitment to providing buyer protection for transactions initiated by agents. Visa CEO Ryan McInerney shared a memorable statistic during the latest earnings call: Visa's average transaction value has decreased from $55 in 2015 to $45 in 2026, but the number of transactions has tripled to 300 billion. Agent payments will push this trend to its extreme—smaller transaction values, higher frequency, and more transactions. This represents a structural growth opportunity for card organizations. Cryptographic native protocols and traditional payment networks are converging from both ends. The final agent payment stack will most likely be a combination of both. From On-Chain to Offline: Agent's Spending Card and Real-World Entry PointFrom On-Chain to Offline: Agent's Spending Card and Real-World Entry PointThe protocol layer solves how Agents pay for APIs. But for the Agent economy to truly take off, there's one last step: enabling Agents to spend money at merchants in the real world.This step saw a breakthrough in May 2026.MoonPay has gone the furthest. The MoonAgents Card, released on May 1st, is a virtual Mastercard debit card. Agents can spend directly from their on-chain self-custodied wallet balance, instantly converting stablecoins to fiat currency. It can be used at any online merchant worldwide that accepts Mastercard. Wallet control remains with the user, and authorization can be revoked at any time. Prior to this card, MoonPay had already built a complete vertical stack: in February, it released MoonPay Agents, integrating CLI, wallet, and autonomous transaction capabilities; in March, it open-sourced Open Wallet Standard, licensed under the MIT license, with participation from PayPal, the Ethereum Foundation, the Solana Foundation, and others. Stripe's Sessions 2026 showcased another approach: Agent-specific virtual debit card issuance services. Enterprises and Agent platforms can programmatically issue one-time virtual debit cards in bulk, each card bound to specific spending rules and spending limits. The advantage of this approach is that it doesn't require stablecoins—it directly utilizes Stripe's existing fiat currency settlement network. The collaboration between Visa and Lightspark opens up another avenue: Visa debit cards supporting stablecoin and Bitcoin balances, covering over 100 countries. On-chain assets serve as the "balance," and the Visa network as the "place to spend." Circle's Agent Stack, launched in May, delves deeper into the underlying infrastructure: the Nanopayments protocol supports gas-free USDC transfers as low as $0.000001, designed specifically for high-frequency machine-to-machine payments. Agent Wallets provide strategically controlled autonomous asset management, and the Agent Marketplace allows agents to discover and purchase services independently. Coupled with its $222 million funding round, Circle's intention is clear: to become a stablecoin infrastructure provider for the agent economy. A particularly vivid example is worth mentioning here. In February of this year, OpenMind's robot dog Bits autonomously walked to a charging station and paid for its charging using USDC nanopayment via Coinbase AgentKit. No humans were present throughout the entire process. A physical agent was using cryptocurrency to purchase energy for itself. MoonPay CEO Ivan Soto-Wright's words can be used to illustrate this point: "Agents are already managing wallets, executing transactions, and transferring value on-chain. The only thing they couldn't do was spend money at merchants. Now they can." The wallet is given to the agent, but who controls the key? The agent can spend money. But who should be responsible for the money the agent spends incorrectly? In January of this year, Step Finance, a DeFi platform on Solana, was attacked. Attackers compromised the devices of the team's executives and gained access to their wallets. What truly amplified the disaster were the AI trading agents integrated into the platform—designed to autonomously execute large transfers without human approval for each transaction. Once the attackers gained access, the agents faithfully performed their designed tasks, only this time transferring money for the attackers. Over 260,000 SOL tokens were transferred, worth nearly $30 million. The platform's token plummeted by 97%, and Step Finance shut down. Every action the agent took was "correct." The problem was that the person who authorized it was no longer the person it was supposed to obey. In April, a security research team from UC Santa Barbara and other universities revealed another dimension of risk: LLM routers sitting between users and AI models. The research documented 26 routers secretly injecting malicious tool calls, one of which directly resulted in $500,000 being siphoned from a customer's encrypted wallet. Attackers don't need to deceive people; they only need to deceive the agent. This is why payment security issues ultimately boil down to the same thing—agent authentication. Can we verify who dispatched an agent? What is it authorized to do? Where are the boundaries of authorization? Who will cover the losses if money is spent incorrectly? Without solving these problems, agent payments can only remain at the level of micro-payments via APIs, amounting to mere cents per transaction—too small to warrant an attack. Once the transaction volume increases to hotel bookings or vault management, the lack of an identity layer becomes unacceptable. Knowing Who Your Agent Is: Four Paths Running Simultaneously Erik Reppel, head of the Coinbase x402 team, aptly put it: "Payments are the 'how' of Agentic Commerce, identity is the 'who's doing it.'" Biometrics: Worldcoin, formerly known as Worldcoin, released AgentKit in March, integrating with x402. Iris verification → World ID → Zero-knowledge proof delegation to the Agent → Server-side verification "This Agent is backed by a real person" but cannot see who it is. The core value is anti-Syrian attack protection—no matter how much money you spend on packages and Agents, the system knows they belong to the same person. Nearly 18 million verified users, but iris scanning faces regulatory hurdles in many countries. KYA Protocol: Skyfire's Know Your Agent issues JWT tokens to each agent—recording the builder, authorizer, scope of authority, and payment method. In late April, Experian, one of the world's three major credit reporting agencies, released the Agent Trust framework, with Cloudflare performing verification at the network edge, Skyfire's KYA providing identity infrastructure, and Experian providing credit assessment capabilities. Card Organization Extension: Visa's Trusted Agent Protocol issues cryptographically signed credentials to agents. Mastercard's Verifiable Intent packages identity, instructions, and transaction results into tamper-proof records. Instead of building an identity system from scratch, it extends the existing credit card trust network to agents. Decentralized Identity on the Chain: ERC-8004 went live on the Ethereum mainnet at the end of January, with each agent being minted as an ERC-721 NFT, growing from 337 registered agents to approximately 130,000. Skyfire has begun mapping KYA identities to ERC-8004 compatible attributes—different systems are intersecting. Each of the four paths addresses a different aspect, and currently none can solve all the problems independently. The banking industry has given this a new name: KYA—Know Your Agent, a natural extension of KYC in the agent era. Where are the opportunities, and where are the pitfalls? At this point, the following clear directions in Agent payments deserve attention: The protocol layer war is almost over, but the application layer war hasn't even begun. x402 has joined the Linux Foundation, and MPP has the backing of Visa and Mastercard. The window for creating another protocol is basically closed. The truly valuable thing going forward is helping traditional enterprises quickly build Agent business scenarios on x402/MPP. Tools, SDKs, and vertical industry solutions—whoever develops them first will reap the rewards. Solana's position needs to be re-evaluated. Google chose Solana as the settlement layer for Pay.sh, 65% of x402 transactions run on Solana, the Solana Foundation actively participates in x402 Foundation governance, and Pay.sh is compatible with both competing protocols. This isn't just another narrative; it's Solana securing a structural position in the Agent economy infrastructure. For long-term investors, this warrants inclusion in their valuation framework. The fundamentals of stablecoins are being rewritten. Previously, demand for stablecoins came from human transactions and cross-border transfers. The Agent economy brings a completely new source of demand—programmatic, continuous, and high-frequency machine-to-machine payments. Stripe spent $1.1 billion to acquire Bridge, and Circle raised $222 million to push its Agent Stack; stablecoin infrastructure is being repriced. If the prediction by Armstrong and CZ, founders of Coinbase and Binance, that "the number of agents will exceed the number of human traders" comes true, the demand curve for USDC will be a structurally upward trend. Identity verification infrastructure is an undervalued sector. World, Skyfire, Visa TAP, Mastercard Verifiable Intent, and ERC-8004—these five solutions each occupy a specific level, and their overlap is only just beginning. Whoever can integrate them into a product that developers can use immediately sits at the intersection of real demand and unmet supply. The security layer is not an add-on feature. Step Finance's $30 million and LLM router's $500,000 are already hefty tuition fees. In a world where agents hold real financial authority, injection vulnerabilities, loss of control, and malicious manipulation are all real risks. ServiceNow specifically demonstrated an "emergency brake switch" for agents at its May conference. Making security a core product feature rather than an afterthought is one of the most important product decisions at this stage. Regulation is the biggest gray rhino. Who is responsible when an agent makes a wrong financial decision? The user? The developer? The model provider? The platform? There is currently no answer. The US GENIUS Act provides a framework for stablecoins, but agent-driven transactions remain a legal gray area. In March, the UK's FCA released a five-level autonomy framework for Agent AI, and the EU AI Act will come into effect in August 2026. Early adopters have the opportunity to define industry standards, while latecomers may be defined by those standards. ConclusionA year ago, in May, x402 was released—an open-source experiment in the crypto world. Twelve months later, AWS, Google, Stripe, Visa, and Mastercard all entered the fray, the Linux Foundation took over standards governance, and a robot dog recharged itself using USDC without human intervention.Behind every node on this timeline stands a company capable of global scale. They're not releasing white papers, they're releasing product announcements. But the pipeline hasn't yet seen truly large-scale commercial traffic flow. The pipes are in place; all that's waiting is water. In the history of technology and business, the distance between the pipes and the water is sometimes shorter than anyone expects—and sometimes longer.