
Visa's Latest Research: Behind 110 Million Transactions, Agent Payments Are Reshaping Internet Commerce
By: Tim Conard (Head of On-Chain Data, Visa), Lucas Shin (Data & Research, Artemis)
Compiled by: Jiahuan, ChainCatcher
Preface
Artemis is an analytics platform in the field of on-chain data, used by institutions such as Visa, Grayscale, Pantera, VanEck, Tether, and Circle to determine what is actually happening on-chain and off-chain. This report uses Artemis Analytics' proprietary dataset, covering real transaction activity on x402 and Machine Payments Protocol, fraud-adjusted metrics, and ecosystem growth.
Visa's goal since inception was not just to build a card network, but to create a global system for the secure exchange of value, independent of specific technologies or forms. Agent commerce is the new frontier on this evolutionary path. As software begins to discover, evaluate, and purchase goods and services on behalf of users and enterprises, new infrastructure will emerge to support trust, authorization, settlement, and control at machine speed.
This new category has the potential to change how payments are initiated, routed, and embedded into digital workflows. This research, completed in collaboration with Artemis Analytics, examines the protocols, standards, data, and market structures currently taking shape around agent payments, aiming to help payment networks, issuers, acquirers, fintech platforms, and enterprise teams understand how this new transaction surface is developing and where the most significant opportunities and challenges are likely to arise.
I. What is Agent Commerce
1.1 Defining Agent Commerce
Agent commerce is an emerging economic form: software capable of continuously and automatically discovering, evaluating, and executing transactions, with transaction scales and operating methods that exceed the original design scope of traditional payment infrastructure.
Autonomy levels vary. Some workflows still retain human review or approval steps in the loop, while others run fully automatically once configured. Agents can purchase SaaS seats, book travel, automatically handle accounts payable, replenish physical inventory, outsource creative work, and often no human intervenes at the moment of ordering.
Within this broad scope, a new subclass has emerged: machine-native, sub-dollar transactions. An AI agent calls an API, receives an HTTP 402 response (Payment Required), evaluates the price, pays, and then consumes the resource. The entire cycle takes milliseconds, requiring no checkout page, no stored cards, and no browsing session—only a wallet or a card, a budget, and a goal.
For payment service providers, agent commerce covers various new transaction types, with economic models, risk control requirements, and infrastructure needs different from traditional e-commerce.
The mark of AI moving from chatbots to autonomous agents is its ability to plan and handle multi-step tasks, call external tools, evaluate results, and autonomously adjust execution strategies. Agents built on frontier models like Claude, GPT, and Gemini are now being deployed in production environments across various business verticals. Over the past year, these agents have become very capable at writing code and conducting research, but they are inherently limited by "whether they can get things."
Empowering agents with payment capability removes this limitation:
- IT agents purchase fifty additional seats on SaaS platforms after personnel changes.
- Admin agents reorder printer toner and pantry supplies when inventory is low.
- Research agents spend a few cents with five financial data providers each for the same company's filing documents, then select the most complete dataset.
- Marketing agents pay 5 cents each to ten image generation service providers to render the same product image, then select the best output.
Agent payments span two distinctly different types of transactions. The first is macro transactions, where agents execute high-value, consumer-like purchases on behalf of humans, such as booking travel, managing subscriptions, and procuring supplies. These transactions are very similar to traditional e-commerce, with the agent acting as a proxy for human intent. The second is micro transactions, characterized by high-frequency, low-value purchases initiated by the agents themselves, such as API calls, data queries, compute access, and tool usage.
Both types exist and are growing, but micro transactions are where new payment architecture is most needed, and where new protocols are filling the gaps.
In the past twelve months, two open-source payment protocols have launched to support the second type of agent payments. x402 (held by the Linux Foundation, incubated by Coinbase and Cloudflare) has processed over $135.7 million in transaction volume since May 2025, with 178.3 million transactions. Machine Payments Protocol (developed by Stripe and Tempo, with contributions from Visa) has settled over $38,000 since going live in mid-March 2026, with approximately 184,600 transactions.
The above figures are raw on-chain cumulative metrics; adjusted figures excluding wash trading and testing are detailed in the methodology description and Chapter IV.
1.2 Why Now
In 1997, online commerce was just getting started. Amazon had been selling books for two years, and PayPal did not exist. When designing the underlying protocols for the commercial internet, engineers reserved the 402 status code (Payment Required), envisioning that future payments could become a native function of the internet, just like loading web pages.
This infrastructure was never built. Card rails carried fixed fees, making sub-dollar commerce unviable; all micro-payment startups in the 2000s failed because the cost of processing a payment exceeded the value of the transaction itself. Since there was no viable way to charge directly for content and services, advertising became the default business model of the internet.
The real change is not just an infrastructure upgrade, but the first appearance of scaled, machine-native payment demand in the market.
The most critical development is AI crossing a capability threshold. Since the release of Claude 4.5 and GPT Codex 5.2 in mid-2025, agents have been able to discover unfamiliar APIs, understand what they offer, evaluate prices, and decide whether to pay, all autonomously and at machine speed, without human approval for each transaction.
This capability 催生 s an organic demand that did not exist before: obtaining APIs, data, compute, and services in a programmatic, on-demand manner. Agents need to pay continuously and in large volumes, without the friction of registering accounts, negotiating contracts in advance, or manually going through checkout processes.
New settlement infrastructure has emerged following this demand signal. A new generation of blockchains has improved the cost and settlement speed of digital payments by several orders of magnitude. Chains built specifically for payments like Tempo can now achieve sub-cent level fees and 500ms settlement. On mature chains like Base and Solana, gas costs have dropped to a fraction of a cent, making micro-payments in the $0.01 to $1.00 range economically viable for the first time.
The key is that this infrastructure itself did not create the demand for agent commerce. Dedicated chains have existed for years without spontaneously growing usage-based API markets. What truly drove market formation is the actual demand generated by increasingly capable AI agents.
Beyond better settlement economics, blockchain rails have also become easier to integrate without exposing underlying complexity: account abstraction, MPC wallets, passkeys, paymaster transactions—increasingly pushing these rails into the background, invisible to end users, merchants, and enterprise teams.
Coinbase launched x402 in May 2025, and Tempo launched MPP in March 2026. Within a few months, both protocols were running real production traffic. Capable agents created the demand, and new protocols and rails made serving this demand cost-effective.
For existing institutions, the risk of inaction is disintermediation. If a few platforms simultaneously control agent payment demand and transaction routing, they can pull payment flows outside existing networks, especially in categories dense with micro-payments where traditional card economic models are weakest. Once this sector forms scale outside the traditional payment stack, existing institutions participating only through high-value agent transactions may miss the highest-frequency part of the emerging agent economy.
Players who can combine protocol-native speed with institutional-level trust, compliance, and distribution capabilities are best positioned to define the next generation of agent payment infrastructure.
II. Transaction Patterns and New Architecture Requirements
2.1 A New Transaction Pattern
Agent payments may appear to have no similarity to human commerce. A person might make a few purchases a day, each深思熟虑 (thoughtful), with merchants they already know. An agent executing a research task might make hundreds of purchases in an hour, each worth a few cents or less, each from a different supplier it just discovered.
Three dimensions define the shape of agent transactions: frequency, amount, and relationship. In human commerce, transactions are low-frequency, medium-to-high value, occurring between parties with existing relationships, such as an account, a subscription, or a stored payment method.
In the micro-transaction category of agent commerce, transactions are high-frequency, low-value, occurring between parties with absolutely no prior relationship. An agent discovers a service, evaluates the price, pays, and leaves. Payment is the only interaction between buyer and seller.
For most of internet history, e-commerce enabled direct payment for physical goods and larger purchases, but the value of digital content and services was often exchanged indirectly. Users paid for content with attention (advertising) or bundled access rights (subscriptions).
Agent commerce makes direct, usage-based exchange possible at any price point; payments can be attached to single queries, single tool calls, single data requests, down to fractions of a cent. In that world, work and payment become inseparable.
As agent commerce scales, transaction volume growth is likely to far outpace value growth. An agent session spending only a few dollars might generate hundreds or even thousands of independent payments; therefore, any system adopting per-transaction authorization mechanisms must complete far more authorization decisions than today when processing the same transaction value.
2.2 Why Agent Micro-Payments Need New Architecture
Card technology stacks were built specifically for human commerce: fewer transactions, larger ticket sizes, pre-established merchant relationships. High-frequency, sub-dollar micro-payments in agent commerce have transaction characteristics 截然不同 (completely different) from traditional card payments, thus requiring new architecture layers on top of the existing card payment system.
Differences manifest in several dimensions. Card rails are optimized for fiat-denominated purchases, second-level authorization, day-level settlement, and manual verification processes. The machine-native micro-payment sector proposes new requirements: sub-cent payment economics, near-instant confirmation, machine-readable credentials, and the ability to transact at high frequency without prior merchant relationships.
Subscription models and metered billing can bypass some of these constraints, but at the cost of requiring account registration, pre-negotiated pricing, and maintaining a continuous business relationship. This model works for known customers but is ineffective for an agent that "discovers a service, pays once, and leaves."

Requirements for Machine-Native Micro-Payments | Source: Artemis, Data as of May 2026
The table above illustrates where machine-native micro-payments require new architecture, but cards are not a rigid, static technology stack. It is flexible infrastructure that can be extended and built upon, which industry leaders are actively exploring.
Visa's Trusted Agent Protocol (TAP), Visa Intelligent Commerce (VIC), and the Visa Card Specification for MPP all demonstrate how card infrastructure can be adapted to support agent-initiated transactions, from higher-value agent purchases to machine-native payment flows.
Therefore, new micro-payment protocols are not intended to replace bank cards, but to cover small-amount transaction scenarios that were previously unserviceable due to high costs on top of the existing card payment system.
This is exactly why new payment protocols are necessary and emerging. They make payment the first and only interaction between buyer and seller, thereby 催生 ing a new class of merchants: APIs, tool servers, developers selling micro-services, data providers charging per query. These businesses need machine-readable payment flows, near-zero onboarding friction, and transaction economics that hold even at extremely small amounts.
III. Deep Dive into Payment Protocols
3.1 Common Foundation
HTTP 402 has been part of the Web specification since 1997. Its design intent was to allow a server (e.g., a webpage) to tell a client (e.g., a user): the resource exists, but payment is required to access it. x402 and MPP are the first two protocols to build production-grade payment infrastructure on top of this status code that has slept for nearly thirty years.
Both protocols follow this basic flow:
- The client sends an HTTP request to the server
- The server returns a 402 status code and a set of payment requirements: price, accepted currencies, where to pay
- The client evaluates the terms, attaches payment, and resubmits the request with credentials
- The server verifies the payment and delivers the resource
The entire exchange happens directly within the HTTP request and response interaction, with no redirect to a payment page, and no independent checkout process or merchant integration. Payment works like authentication on the Web today, via headers in the HTTP request. Any server that returns a 402 can become a paid endpoint, meaning any agent with money can pay for it natively.
3.2 MPP: Two-Party Settlement
Trust Model. MPP has no intermediary. The agent pays the server directly, and the server delivers the resource.
Settlement. The main settlement rail is Tempo, a blockchain built specifically for payments, with finality of approximately 500ms and fees below one-tenth of a cent. Tempo also supports gasless transactions, meaning the agent or server can cover gas fees for the other party, reducing friction for new participants.
MPP also natively supports cards through Visa's Card Specification and Stripe's integration, and supports the Lightning Network within the same protocol. A server can offer multiple payment options simultaneously, and the agent chooses any rail it prefers using the same protocol flow. This design pushes more value to the platform, tool, and trust layers above the transaction.
Payment Types. MPP currently supports two payment methods. The first is charge, a one-time payment where the agent pays and the server delivers. A single charge can be atomically split among up to ten recipients within one transaction, enabling marketplace models and revenue sharing at the protocol level.
The second is session, a streaming payment channel where the agent deposits funds into on-chain escrow, and the server delivers value incrementally. The agent issues off-chain credentials during service consumption, authorizing the server to claim funds corresponding to delivered value. The server can verify these credentials in microseconds without initiating an on-chain call for every request, allowing session operation speeds to approach a standard API call. Unconsumed funds can be refunded to the agent, and the server settles credentials in batch at the end of the session.
Governance. MPP is developed on the IETF standards track (draft-httpauth-payment). The IETF (Internet Engineering Task Force) is the body responsible for driving most core technical standards of the internet, including HTTP itself, TLS (encryption behind HTTPS), and DNS. Entering the IETF standards track means the protocol undergoes a formal, open review process where any engineer or organization can participate, comment, and propose modifications. This makes MPP the only agent payment protocol developed through "the same process that produced internet foundational protocols." Specifications are publicly available, and design remains vendor-neutral.
3.3 x402: Three-Party Settlement
Trust Model. x402 introduces a third party in the exchange, called a facilitator. When an agent pays for a resource via x402, funds go to the facilitator first, which holds the funds, verifies whether the server delivered what was promised, and releases the money to the seller only after confirmation.
Settlement. x402 is an open payment protocol designed to run on any chain where a facilitator is deployed, not limited to a fixed set of blockchains. In practice, today's activity is concentrated on Base, Solana, and Polygon, with Base carrying the vast majority of transaction volume. Payments on these chains are settled on-chain in stablecoins, primarily USDC.
The facilitator architecture itself is also rail-agnostic. Although current facilitators mainly settle on-chain, this architecture is equally compatible with facilitators settling via traditional rails such as ACH, SEPA, or card networks. This structure makes the facilitator the key monetization and trust layer in the x402 model.
Payment Types. x402 currently supports standard fixed-price payments, and a usage-based billing mode called Upto. With Upto, the agent authorizes a cap amount and pays only for the portion actually consumed. This is designed for variable loads where the agent does not know in advance how much resource it will use.
Governance. x402 was originally created by Coinbase as a proprietary specification. In April 2026, custody of the protocol was transferred to the Linux Foundation, which established the x402 Foundation as a vendor-neutral governance body. The x402 Foundation was initially driven jointly by Coinbase, Cloudflare, and Stripe, and launched with support from Visa, Amazon Web Services, American Express, Base, Circle, Fiserv Merchant Solutions, Google, KakaoPay, Mastercard, Merit Systems, Microsoft, Polygon Labs, Shopify, and the Solana Foundation.
Moving to open governance under the Linux Foundation signals the intent to solidify x402 as a shared industry standard rather than a single company's project.
3.4 What These Differences Mean
Both protocols are very new and evolving. The chains, payment methods, currencies, and features described above reflect the status currently online.
Both architectures are designed to be scalable: the rail-agnostic facilitator model of x402 means accessing a new facilitator adds any new settlement rail; the multi-rail design of MPP means new payment methods can be supported at the protocol level. As the two protocols mature, the feature set will continue to expand.
IV. What the Data Shows
Unless otherwise noted, all metrics in this chapter are adjusted according to Artemis Analytics' proprietary methodology.
4.1 x402: 11 Months of Real-Time Data

x402 Cumulative Transaction Volume (Adjusted), October 20, 2025 to April 21, 2026 | Source: Artemis
Since launching in May 2025, x402 has processed $15 million in adjusted transaction volume and 109.6 million adjusted transactions. Activity accelerated sharply in October 2025, with monthly transaction volume rising from about 40,000 to 3.8 million. In November 2025 alone, the protocol processed approximately 38 million transactions.

x402 Average Transaction Amount (Adjusted) | Source: Artemis
Transaction volume and value diverged significantly over time. November 2025 was the peak month by transaction count, with 38 million transactions generating $5.15 million in adjusted transaction volume. March 2026 presented a completely different pattern, with $1.64 million in adjusted transaction volume corresponding to only 2.1 million transactions, the highest average per transaction on record.
The top 1% of buyers (approximately 4,000 wallets) accounted for about 90% of adjusted transaction volume and about 50% of adjusted transaction count. This concentration pushed up the average transaction amount. However, the broader trend still indicates that the payment scale and workflow combinations supported by x402 are wider than the term "micro-payments" itself implies.

x402 Transaction Count by Chain (Adjusted) | Source: Artemis
On-Chain Distribution. Base overwhelmingly dominates this ecosystem, accounting for about 90% of adjusted transaction count and 93% of adjusted transaction volume. Solana ranks second by transaction count, and Polygon third. At this stage, x402 is clearly a Base-centric market, with other chains contributing meaningful but secondary activity.

x402 Transaction Count Share by Category (Adjusted) | Source: Artemis
Category Distribution. By transaction count, Agent-to-Agent Services is the far-and-away largest category, reflecting exactly the high-frequency, low-value transaction pattern that defines machine-native commerce.
This category captures chained workflows: one agent pays another agent for a sub-task, such as data retrieval, task execution, or inference, meaning a single user request can generate a chain of transactions. Unclassified long-tail endpoints account for most of the transaction volume, highlighting fragmentation on the supply side and the need for better discovery and aggregation tools.

x402 Buyer Spending Concentration (Adjusted) | Source: Artemis
Buyer Concentration. Dollar spending is highly concentrated in a very narrow set of buyers. Less than 0.02% of buyers generated about 48% of adjusted transaction volume. This is consistent with the characteristics of an early market: even as the overall participant base expands, a relatively small number of mature participants still drive most economic activity. To date, x402 has over 422,000 adjusted buyers.

x402 Cumulative Seller Count (Adjusted) | Source: Artemis
Supply Side. x402 has recorded approximately 5,300 adjusted sellers (i.e., merchants) to date. Traditional merchant onboarding often takes weeks and requires a payment processor, whereas on x402, any developer can deploy a collectible endpoint and start accepting payments immediately.
Two tailwinds suggest the supply side of agent-consumable resources will continue to expand: First, AI is lowering the barrier to creating digital products, meaning more people and agents will become merchants; Second, new Web services and applications are increasingly designed for agents, API-first, instantly accessible, and priced per request.
4.2 Machine Payments Protocol: First 33 Days
MPP launched in mid-March 2026 and had been online for just over a month at the time of this analysis. Given the protocol is still in an early stage, granular breakdowns are less meaningful than for x402; the focus here is on total activity and the structural differences emerging between MPP's two payment types.

MPP Cumulative Transaction Count (Adjusted), March 18 to April 21, 2026 | Source: Artemis
Since launch, MPP has processed approximately $25,000 in adjusted transaction volume and about 115,000 adjusted transactions. Transaction growth has been stable and roughly linear from day one, averaging about 4,000 adjusted transactions per day.

MPP Transaction Volume by Payment Type (Adjusted), charge vs session | Source: Artemis
charge accounts for 97% of adjusted transaction count. session amounts per transaction are significantly larger but remain a very small part of overall activity. As described in Section 3.2, session allows agents to deposit funds into on-chain escrow channels and pay incrementally during service consumption. Since servers settle accumulated credentials in batch rather than per-request, even if underlying usage is granular, a single session transaction appears larger than charge. Early session activity indicates this model is being tested but has not yet gained widespread adoption.

MPP Average Transaction Amount by Payment Type (Adjusted) | Source: Artemis
Transaction volume accumulates gradually, with most transactions still in the sub-cent range. The average transaction amount for charge remained steady around $0.17 during this period, while the average for session was $2.15. The upward trend in session amounts likely reflects developers using larger deposit amounts for experimentation while testing the streaming payment model.

MPP Cumulative Buyer and Seller Count (Adjusted) | Source: Artemis
MPP has recorded over 2,800 adjusted buyers and 90 adjusted sellers to date. The approximately 32:1 buyer-to-seller ratio indicates demand is running ahead of supply, with more agents transacting on a still limited set of available endpoints. The supply side remains underdeveloped. x402 saw seller counts of a similar magnitude in its first few months before accelerating sharply.
V. Trust, Identity, Spending Control, and Dispute Resolution
5.1 Agent Trust is the Hardest Problem
Traditional commerce defaults to the buyer being a human, possessing intent, judgment, and accountability. Agent commerce challenges all three. The buyer is software acting based on authorization, making decisions at machine speed, often without real-time human supervision. This creates a trust problem that no protocol can fully solve at the protocol layer alone.
Risk of Buying Wrong. The simplest failure mode is the agent buying the wrong thing. The agent misunderstands the task, selects the wrong service, or pays too high a price for a resource. In human commerce, the buyer discovers errors before checkout or initiates a dispute afterwards.
Autonomously running agents may not even realize this error, and at machine speed, a misconfigured agent can execute hundreds of bad transactions before anyone notices, creating unmanageable reconciliation, refund, and dispute burdens at scale.
Adversarial Attacks. Prompt injection, where malicious input manipulates the agent into executing unauthorized actions, remains an active area of research without a complete solution. Attackers who breach the agent's reasoning process can redirect purchases, steal funds, or trigger unauthorized spending across multiple connected services. The attack surface grows with autonomy: the higher the degree of freedom in agent transactions, the greater the damage a compromised agent can cause.
Liability Attribution. When an agent makes an unauthorized purchase, the liable party may not be clear initially. The person who granted permission, the platform hosting the agent, the model provider whose reasoning failed, the merchant who accepted the payment—all parties have tenable claims and defenses. Existing legal and regulatory frameworks were not designed for this chain of delegation, and clear precedents may not exist.
Cascading Failures. As agent commerce matures, agents will increasingly transact with other agents rather than directly with merchants. A compliant agent might pay a document review agent to analyze a contract, which then pays a legal research agent to retrieve relevant case law, which then pays a database agent to fetch court documents.
These agent-to-agent chains create value through specialization but introduce risk through interdependence. If any agent in the chain delivers poor results, every agent upstream has already paid for work based on that output. Correcting this error requires propagating the dispute back along the chain hop by hop. This multi-hop payment chain has no counterpart in traditional commerce, nor do existing mechanisms exist to unwind them.
5.2 Current Solutions
The trust challenges above are being addressed at multiple layers of the technology stack, with players from both crypto-native and card-native sides moving.
Protocol Layer. Payment protocols like x402 and MPP are building trust directly into the transaction mechanism. The third-party facilitator approach can verify delivery before releasing funds, escrow mechanisms allow buyers to pay only for consumed value, and usage-based authorization limits exposure by capping spending per single request. These are protocol-level protections, but their effectiveness depends on which protocol and payment type buyers and sellers choose to use.
Identity Layer. The ecosystem remains fragmented. Most agents use API keys or wallet addresses for authentication, which verifies access rights but says nothing about history or reliability.
Public registries like x402scan and MPPscan provide basic endpoint discovery. Coinbase's Bazaar registry maps merchants to收款 addresses. ERC-8004 (Trustless Agents Standard) aims to make agent reputation portable across networks, allowing merchants to verify an agent's transaction history before deciding to transact.
On the card side, Visa's TAP, Mastercard's Agent Pay, and Stripe and OpenAI's ACP help anchor agent identity to existing authorization frameworks, inheriting decades of accumulated KYC and risk control infrastructure, but all are tied to card rails.
Policy Layer. Programmable spending control is emerging simultaneously in both ecosystems. Coinbase Agentic Wallets provide paymaster transactions with programmable limits on Base, while policy layers like Turnkey, Privy, and Safe encode whitelists, spending caps, and merchant routing directly into wallet infrastructure. On the card side, tokenized credentials with issuer-defined spending rules serve the same function through familiar infrastructure.
5.3 Unresolved Issues
Currently, there is no universal agent identity standard that can connect emerging crypto-native rails with existing traditional rails. Reputation built in one ecosystem cannot be transferred to another.
Disputes lack existing handling mechanisms. Chargeback windows and evidence requirements were originally designed for human-speed commerce with clear buyer intent. When an agent executes thousands of transactions per hour, the concept of a "single disputed order" cannot map to hundreds of sub-cent API calls.
Furthermore, liability remains legally ambiguous. No jurisdiction has established clear precedents specifying how liability should be allocated among the person granting permission, the platform hosting the agent, the model provider, and the merchant.
These gaps represent both the most significant risks in agent commerce today and the clearest value-add opportunities for existing institutions in the traditional payment ecosystem.
VI. Standards Landscape
6.1 Agent Commerce Technology Stack
Agent commerce requires coordination across multiple infrastructure layers, from how agents communicate to how money moves. No single standard covers the complete technology stack, and this landscape is evolving rapidly.

Agent Commerce Standards Landscape, Protocols Emerging at Each Layer | Source: Artemis, Data as of May 2026
Together, these standards constitute the infrastructure stack for agent commerce taking shape, from communication and identity to authorization and payment execution.
6.2 Card-Native Protocols
x402 and MPP define how agents pay for machine-native resources; in parallel, another batch of protocols is emerging for scenarios where "agents complete transactions on behalf of human consumers through existing card infrastructure."
Visa TAP. Trusted Agent Protocol allows merchants to verify via cryptographically signed HTTP messages that an AI agent is indeed authorized to act on behalf of a user. TAP adds a standardized trust layer on top of existing Web and merchant infrastructure, allowing merchants to handle agent-initiated transactions without major changes to their payment stack. TAP was co-created with Cloudflare, with input from Shopify, Microsoft, and Stripe, launched in October 2025, with early adopters including Nuvei, Adyen, and Stripe. For the traditional payment ecosystem, TAP is the most direct path from existing card infrastructure to agent-initiated commerce.
ACP (Stripe + OpenAI). Agentic Commerce Protocol allows agents to complete purchases through existing merchant checkout processes using card payments. Agent platforms and merchants must pre-approve each other for transactions to occur, forming a curated market where only audited participants can trade. Payments are processed by Stripe, and OpenAI charges merchants a fee on top of standard processing costs.
UCP (Google/Gemini + Shopify). Universal Checkout Protocol allows agents to complete purchases directly within Google's product surfaces, such as Search AI Mode and the Gemini app. Merchants retain full ownership of the customer relationship and remain the merchant of record in all transactions. The protocol was jointly developed by Google and Shopify, with over 20 participants including Etsy, Wayfair, Target, and Walmart, supporting REST, MCP, AP2, and A2A integrations.
VIC (Visa). Visa Intelligent Commerce is a single integration point released in April 2026, allowing merchants, acquirers, and agent platforms to participate in agent commerce. This platform is protocol-agnostic, supporting TAP, x402, MPP, ACP, and UCP through a single integration. It provides tokenization, passkey-based authentication, and programmable spending control, allowing users to set amount limits, merchant categories, or require real-time approval. For merchants, VIC eliminates the need to integrate each agent protocol individually. For the payment ecosystem, it positions Visa as the connection layer between agents and existing commercial infrastructure, with over 100 partners to date, 30 of whom are actively building in the VIC sandbox.
AP2 (Google + 60 Partners). Agent Payments Protocol introduces cryptographically signed mandates as the basis for agent authorization. Intent Mandates record the conditions under which an agent can purchase on behalf of a user, Cart Mandates record specific items and prices for a transaction, and Payment Mandates authorize the final payment. The protocol is payment rail-agnostic, supporting cards, bank transfers, and cryptocurrency. AP2 is designed as an extension of both A2A and MCP, positioning itself as an authorization layer that can sit atop various payment protocols. Partners include Mastercard, American Express, PayPal, Adyen, Coinbase, and Salesforce.
Agent Pay (Mastercard). Agent Pay uses Mastercard's tokenization technology to create agent-specific payment credentials. Consumers authorize agents to act on their behalf within set parameters, and agents complete purchases using the tokenized version of the consumer's card. Spending limits, merchant categories, and other guardrails are programmable and bound directly to the credential.
6.3 Convergence and Conditions for Scale
The line between crypto-native protocols and card-native protocols has begun to blur. MPP now covers both on-chain crypto payments and fiat payments via shared payment tokens, and Visa's Card Specification SDK is designed to extend the protocol into card-based agent commerce. The practical effect is that a single machine payment framework can increasingly support both stablecoin flows and card transactions.
For card networks and issuers, this opens a path into agent payment flows without requiring merchants to adopt a completely separate technology stack. From the other direction, Stripe added support for x402 payments in USDC on Base in March 2026, connecting stablecoin-native payment flows into the broader agent commerce ecosystem.
The direction points to convergence rather than competition: bank cards are better suited for agents completing traditional consumption on behalf of humans, stablecoins are better for machine-native micro-payments, and hybrid workflows may use both rails simultaneously.
For this convergence to move beyond the early adopter phase, three things need to happen:
- Interoperability between standards needs to mature. Over time, agents should be able to combine AP2-style authorization mandates, payment execution via x402 or MPP, and trust and card-end verification via TAP or similar systems within a single workflow, without requiring custom integration for every combination.
- Regulatory frameworks need to catch up with technology. Agent-initiated transactions, liability attribution, and how to handle cross-border agent commerce remain undecided, especially when these systems begin operating across multiple payment rails and jurisdictions.
- Existing institutions need clear roles. These institutions, which process most of the world's payments, need to see their position in the technology stack: as issuers of agent credentials, processors of agent transactions, or providers of trust and dispute infrastructure currently lacking in crypto-native protocols.
VII. Use Cases
7.1 Overview
The highest volume use cases today are lead enrichment, data scraping, inference and inference routing, and web search. Lead enrichment is one obvious high-value workflow, where agents pay per query to pull individual or company profiles, verify emails, and complete CRM data.
Beyond this, a growing long tail of services is emerging: social media intelligence (scraping LinkedIn, X, and Reddit data), image and video generation endpoints, Google Maps queries, phone call APIs, and agentic search.
One example illustrates where this is heading: face search endpoints, where an agent uploads an image and receives the highest probability identity match results across the web. Such niche but high-value services are never worth the overhead of a full SaaS product, but make perfect sense as a per-request收费 API that agents can discover and call. As the ecosystem matures, such creative long-tail scenarios are likely to emerge in large numbers.

What Agents Are Paying For, Early Categories of Machine-Initiated Spending | Source: Artemis, Data as of May 2026
7.2 Per-Query Services
A research agent building a company profile might spend a few cents with multiple data providers each to query, compare results, and select the best one. Total cost might be less than a dollar.
In the traditional model, the same workflow often requires subscribing to a single supplier and lacks the ability to compare prices at the moment of need. This is one of the clearest business forms released by low-friction payments. The agent has no account, no pre-negotiated contract, and no long-term relationship with the seller. It discovers the endpoint, pays, evaluates the output, and leaves.
The applicability of this model extends far beyond data completion. Developers can expose document parsers, translation services, code review tools, or other MCP-accessible functions as paid endpoints. Agents can discover and call these tools on demand, and providers do not need to build an entire SaaS business around them. This lowers the barrier to becoming a merchant in the machine economy and 催生 s a long tail of specialized, per-use paid services.
This model has already begun to be productized. AgentCash built by Merit Systems packages wallet management, merchant discovery, and access to 420+ paid APIs into a single integration for agent frameworks like Claude Code. This platform allows agents to access hundreds of paid services without individual configuration, abstracting away the complexity of protocol-layer payments on x402 and MPP.
7.3 Streaming and Metered Services
Not all agent payments are single request-response transactions. Some workflows require continuous access to a resource over a period of time. A coding agent might consume compute continuously for hours, a monitoring agent might continuously receive market data, paying only for the portion of the data stream it actually uses.
MPP's session is designed for this model. The agent deposits funds, pays incrementally as it consumes, and can stop at any time when the task ends or quality falls below expectations. Merchants are compensated for delivered usage, and buyers do not need to commit to a fixed subscription for scenarios with fluctuating demand. This model is more suitable for agent workflows with continuous, unpredictable usage tightly bound to task completion.
7.4 Agent-to-Agent Services
A signature scenario in agent commerce is agents transacting directly with other agents. A single user request can trigger a chain of machine-to-machine purchases, with one agent delegating sub-tasks to specialized services: search, data completion, compute, execution, each paid programmatically at the moment of need.
This is a completely new commercial surface. There are no checkout pages, no shopping carts, and no human approval steps between purchases. The workflow runs continuously, with each agent evaluating and paying for the next service in the chain. As described in Chapter IV, this is already the largest activity category by transaction count on x402.
This is one of the clearest examples of commerce created by agents themselves. Payments are embedded directly into the execution of work, allowing specialized services to be coordinated and traded in real time.
VIII. Economic Models and Value Attribution
8.1 Monetization Models
The scenarios above share a common characteristic: high frequency, low value, and no pre-existing relationship between buyer and seller. The economic model for serving this commerce is very different from traditional payments, and the two payment standard ecosystems (x402 and MPP) currently capture value at different positions.

Value Attribution in Agent Commerce, Comparison of x402 and MPP Ecosystems | Source: Artemis, Data as of May 2026
In x402, transaction volume flows through the facilitator layer, where intermediaries can charge processing fees for verification, settlement, and related trust functions. In MPP, less value is attributed at the transaction layer, with more归属 ing to the surrounding ecosystem: platforms aggregating demand, tools simplifying integration, and trust or discovery layers helping buyers and sellers transact with confidence.
This difference shapes how the two ecosystems expand. x402 monetizes the intermediary role in the fund flow, while MPP pushes the monetization focus towards distribution, routing, and trust. For existing institutions, the key question is where sustainable economics will ultimately land.
8.2 Sustainable Advantages
Agent commerce makes it exceptionally easy for new merchants to enter the market. Developers can put an existing API behind a 402-compatible paywall with relatively small integration overhead; the same basic logic applies to the broader machine payment ecosystem. This is by design. x402 and MPP were built to lower supply-side onboarding friction, making it easier for new services to monetize.
Easy entry does not equal sustainable economics. When new endpoints can be launched quickly and competitive supply is easily replicated, value shifts from individual merchants to layers that are harder to replicate. In e-commerce, similar dynamics pushed sustainable value towards platforms and payment infrastructure rather than the merchant long tail they empower; in agent commerce, the strongest structural advantages sit in distribution, trust, and infrastructure.
Whoever controls how agents discover and route to services controls demand. Whoever builds identity and reputation systems helping agents choose between competing suppliers reduces fraud and increases conversion. And whoever provides settlement, compliance, and dispute infrastructure is embedded in the transaction flow itself.
This is exactly where the traditional payment industry has natural advantages. The capabilities payment networks, issuers, and acquirers spent decades building are precisely what is still needed when agent commerce scales: identity verification, spending policy enforcement, fraud detection, cross-border settlement, and merchant risk assessment.
New protocols provide programmability, speed, and cost structures traditional card rails cannot match on extremely small transactions. But they do not currently offer trust infrastructure, regulatory coverage, or merchant distribution capabilities at the scale of existing institutions.
IX. Compliance and Regulatory Considerations
9.1 How These Protocols Map to Existing Regulations
Agent payment protocols do not map cleanly onto existing payment regulations.
In an x402-style model, the facilitator may temporarily hold funds, verify delivery conditions, and then release funds to the seller. Depending on how this role is designed and how a jurisdiction interprets it, it may touch money transmission, payment processing, escrow, or related licensing frameworks. An MPP-style direct connection model reduces the role of dedicated intermediaries but raises another question: when no party is clearly between buyer and seller, who bears the regulatory obligations related to the transaction?
Cross-border complexity adds another layer of uncertainty. An agent within one jurisdiction may pay a server in another jurisdiction, while infrastructure or service providers involve even more jurisdictions. Existing cross-border payment frameworks were built on the premise that counterparties are identifiable, jurisdictional归属 is clear, and transaction flows are slow, whereas agent transactions compress these premises. For payment service providers, this means compliance architecture may evolve from a back-office function into a source of competitive differentiation.
9.2 Stablecoin Compliance
The early settlement paths of both protocols directly link stablecoin regulation to the future of agent commerce. In the US, the GENIUS Act established a federal licensing and reserve framework for payment stablecoin issuers, providing a clearer regulatory foundation for stablecoin-based payment flows than before.
For x402, activity has been closely tied to USDC, especially on Base, giving this ecosystem a relatively institution-friendly starting point. But x402 itself is not necessarily limited to a single stablecoin or single chain; specific implementations can vary with facilitators and networks.
For MPP, current implementations include crypto-native settlement mechanisms like pathUSD on Tempo, but the protocol's scope is wider than any single rail and is also designed to support fiat-pegged payment methods like shared payment tokens.
For traditional payment audiences, stablecoin compliance matters because it determines whether these fund flows can be integrated into regulated financial infrastructure. If stablecoin settlement has a clear regulatory status, it becomes a more viable rail for banks, acquirers, and processors to build upon. If regulatory treatment remains fragmented across jurisdictions, institutional adoption will slow down.
9.3 Agent-Specific Considerations
Existing KYC and AML frameworks are designed around human customers. In agent commerce, this premise fails.
The core compliance question becomes: for purposes of identification, monitoring, and supervision, who is the truly relevant "customer"? Is it the person granting permission to the agent, the platform operating the agent, the entity funding the wallet, or some combination of the three? Regulatory treatment on this point remains undecided.
Tax and reporting obligations face similar challenges. An agent may execute thousands of small transactions across multiple jurisdictions in a single session, generating reporting burdens existing systems were not designed to carry. Buyers and sellers may have only limited visibility into each other's identity, location, or legal status.
For payment service providers evaluating this market, the infrastructure required to make agent transactions traceable, auditable, and reportable at scale is itself likely to become a significant value-creation layer.
X. Outlook
10.1 Future Opportunities
Agent commerce has begun to operate but remains in an early development stage overall. Protocols exist, transaction patterns are visible, while surrounding infrastructure remains underbuilt. This creates several clear growth and investment opportunities:
Hybrid Settlement of Cards and Stablecoins. For issuers and acquirers, one of the most direct opportunities is hybrid settlement: cards on the front end, stablecoins on the back end. Cardholder experience remains unchanged, while settlement shifts to stablecoin rails, supporting faster receipt of funds, continuous settlement windows, and lower cross-border costs.
This gives existing institutions a way to share the speed and cost advantages of stablecoin infrastructure without requiring merchants or consumers to change behavior. As agent commerce scales, hybrid settlement is likely to become an attractive path for institutions wishing to access machine-initiated payment flows through infrastructure they already operate.
Protocol Convergence and Interoperability. x402 and MPP currently operate as independent ecosystems with different settlement rails, trust models, and merchant bases. Over time, agents need to be able to route transactions between protocols without custom integration for each one.
Early signs of convergence are already appearing: MPP supports both crypto and card-linked payment flows, while Stripe's support for x402 on Base connects stablecoin-native payments into broader payment infrastructure. Interoperability layers helping agents choose the correct protocol for each transaction type could become a significant control point in the technology stack.
Cross-Border Agent Commerce. Cross-border activity is likely to be a default feature of agent commerce, not a niche edge case. Agents can initiate transactions across jurisdictions, bypassing traditional identity and residence anchors relied upon by existing compliance systems. Infrastructure for handling compliance, currency conversion, and jurisdictional routing at machine speed remains immature. For payment networks with global coverage, this is a natural extension of existing capabilities and one of the clearest differentiation value opportunities in the near term.
Agents as Merchants. Today agents primarily act as buyers, but over time, they will also become sellers. As the cost of building and deploying software continues to decline, more developers, companies, and workflows will have the capability to expose specialized services directly to agents. This will significantly expand the ecosystem's supply side, bringing more agent-accessible endpoints, more niche services, and more transaction volume flowing through the payment rails supporting them.
10.2 Conclusion
Agent commerce is taking shape. x402 and Machine Payments Protocol are processing real transactions, merchants and agent platforms are beginning to build on top of them, and the surrounding standards landscape is also converging around a set of common needs: trust, identity, authorization, settlement.
The transaction patterns emerging in this market are high-frequency, sub-dollar, autonomous, and often cross-border. Protocols solve for speed, programmability, and cost. But to scale into the regulated financial system, it still requires what existing payment players spent decades building: trust infrastructure, compliance frameworks, merchant distribution, and dispute resolution.
This gap defines the opportunity, and Visa has already positioned itself on both sides of the market. TAP and Visa Intelligent Commerce connect card infrastructure to higher-value agent transactions, while the Card Specification SDK for MPP and participation in the x402 Foundation extend Visa's reach into machine-native micro-payments.
This breadth of coverage reflects exactly the convergence judgment throughout this piece: macro and micro transaction models may start in independent ecosystems, but over time, they are likely to connect through shared trust, identity, and settlement layers.
This market is still in an early stage, with many key infrastructures yet to be finalized. Institutions that participate now in building trust mechanisms, spending controls, and protocol interoperability are most likely to continue occupying core positions after agent commerce moves to scale.















