When agents become consumers, who will rewrite the underlying logic of internet commerce?

Author: YQ

Compiled by: Jiahua, ChainCatcher

The business model of the internet is advertising. For thirty years, this has become the default rule: show content to humans, harvest attention, and convert clicks into revenue. Search engines, social networks, news websites, and video platforms all follow the same logic without exception. Users are not customers. Users are products.

Agents have broken this model. Agents that call APIs do not have attention to harvest. They do not view banner ads. They do not click sponsored links. They do not make impulse purchases based on influencer recommendations. They pay for services based on utility assessment; otherwise, they turn away. The entire advertising economy is built on the assumption that there are humans on the other side of the screen. When users become machines, this assumption collapses.

This is not an assumption. Gartner predicts that by 2028, $15 trillion of B2B procurement will be completed through AI Agents. ChatGPT has 900 million weekly active users (data from February 2026). When these users delegate purchasing decisions to agents, the agents need a payment method. Two companies are paving this path: Coinbase's x402 and Stripe's MPP. Both have activated the same status code that has been dormant in the HTTP specification since 1996. Beyond this, they cannot reach a consensus on anything else.

Awakening the Sleeping Payment Protocol

In 1996, Roy Fielding embedded a placeholder in HTTP/1.1: status code 402, "Payment Required." The vision at that time was to embed micropayments (very small financial transactions) at the network's core. Reading an article would cost five cents. Loading an image would cost one cent. Making money a native part of HTTP, like links.

Three things killed it. Credit card fees of $0.25 to $0.35 per transaction made five-cent payments absurd. Payment prompts caused decision fatigue. There were no browser wallets at that time. Microsoft's MSN micropayment failed. DigiCash failed. The internet turned to advertising, and HTTP 402 lay dormant for thirty years.

What has changed is not the technology. Stablecoins and Layer 2 networks have made transactions below one cent possible. What has changed is the users. When users are human, advertising works. When users are machines, advertising becomes structurally impossible. The machine economy needs a payment layer. HTTP 402 is clearly an excellent place to build it. Coinbase launched x402 in May 2025. Stripe launched MPP in March 2026. The race has begun.

But everyone is asking which protocol will win. This is the wrong question. The payment track is just infrastructure. Visa processes $14 trillion in transactions annually. Mastercard and Santander completed the first AI Agents payment in Europe on March 2, 2026. When 4 billion existing bank cards can be used for agent transactions, the protocol for transferring funds is just a ticket to entry. The real moat is the decision orchestration layer: in this system, agents decide what to buy, from whom to buy, and how to verify trust.

Trillion-Dollar Blue Ocean

In the past twelve months, five analytical firms have released forecasts for the Agents business. They differ in specific numbers. But they are surprisingly consistent in the direction of development.

Morgan Stanley's lower limit ($190 billion) and Gartner's B2B data ($15 trillion) differ by two orders of magnitude. This gap is not due to confusion but rather different statistical scopes. Morgan Stanley calculates U.S. consumer e-commerce. Gartner calculates all global B2B procurement. Even the most conservative bottom-line estimate represents a market size larger than Stripe's total processing volume for all of 2023 ($1 trillion).

Why Enterprise Applications Must Lead

Bain & Company surveyed 2,016 U.S. consumers in early 2026. For anyone predicting explosive growth curves, these numbers are undoubtedly a cold shower.

The key data is: only 7% of people trust third-party AI platforms to shop for them. The trust level for retail chatbots is 25%. The gap here is the distance of trust. People trust stores they have already shopped at more than a generic AI agent. The first wave of agent business will emerge through branded retail AI, rather than through independent agent protocols.

The subtext for x402 and MPP is the same. Both protocols need to gain enterprise adoption before consumer adoption becomes widespread. Consumers do not choose the payment track; merchants do. And merchants choose the track based on fraud rates, settlement speed, and chargeback liability, rather than the elegance of the protocol.

Composable Payment Protocol Ecosystem

The fastest way to misunderstand agent payments is to list every protocol's announcement side by side and then ask, "Which is best?" These protocols occupy different levels of the same tech stack. Mixing them up is like comparing the TCP protocol with Shopify.

x402 is at the bottom layer. It adds payment semantics to HTTP itself. Any server can return a 402 status code, and any client can respond with a signed transaction. No SDK, no session, no state required. MPP is one layer up. It uses the same 402 status code but wraps it in a session protocol that aggregates micropayments, supports fiat currencies, and connects to Stripe's compliance framework.

Above these two is the decision orchestration layer. Google's AP2 defines how agents discover each other, exchange authorizations, and delegate permissions. The ACP, built in collaboration with OpenAI, is responsible for product discovery, cart management, and checkout. Visa and Mastercard build the credential layer on top of all this.

Most analyses overlook one insight: these are composable layers, not exclusive choices. In theory, a single transaction can be authorized through AP2, product selection through ACP, session management through MPP, and finally settled through x402. No one has yet built a complete stack. This composability is a design assumption rather than a production reality. But its architectural logic is correct: Google's a2a-x402 extension (open-sourced on GitHub) has already bridged AP2 and x402.

The question now is, which layer can capture the most value as other layers become homogenized.

Costs Determine Winners and Losers

Every discussion about agent payments ultimately turns into a discussion about costs. Here are actual data comparisons of four proposals that others have yet to publish.

x402 Protocol: Minimal, Open Source, and Pure Crypto Route

The x402 protocol can be explained with five lines of middleware code. The client requests a resource. The server returns an HTTP 402 with a payment request header. The client signs a stablecoin transfer and retries with the payment signature. The service provider verifies the settlement on-chain. The server returns HTTP 200.

As of March 2026, on-chain data shows: an average daily transaction volume of about 131,000, an average daily transaction amount of about $28, and an average payment amount of $0.20. According to CoinDesk's forensic analysis (March 11, 2026), about half of these are developer tests and integration probes. The annualized total transaction volume is about $10 million. After excluding tests, the annualized real commercial transaction volume is about $5 million. For a protocol supported by Coinbase, Cloudflare, Visa, and Google, this is just seed round data. The protocol itself operates well; the market just hasn't arrived yet.

What makes x402 architecturally powerful is precisely what it lacks. Zero protocol fees. Zero account creation. Zero vendor lock-in. The specification adopts the Apache 2.0 license. Any server in any language can be deployed in an afternoon. The x402 Foundation currently has Coinbase, Cloudflare, Google, and Visa as members, with AWS, Circle, Anthropic, and Vercel as additional partners. Its trial-and-error cost is essentially zero.

The v2 specification addresses the original protocol's obvious limitations. The plugin architecture allows service providers to support multiple blockchains. Reusable sessions reduce on-chain overhead for each request. Multi-chain support has expanded from the Base network to the Ethereum mainnet, Arbitrum, and Solana.

Its weakness is structural. x402 only supports cryptocurrencies. There is no fiat currency channel. If an agent needs to pay for a SaaS API with a corporate credit card, it cannot use x402. Moreover, the underlying Coinbase platform is currently undergoing a transformation.

Stripe MPP Protocol: Anchoring Compliance and Fiat Payments

x402 lacks a fiat currency channel. It cannot aggregate micropayments. It has no built-in fraud detection, no tax calculation, and no refund logic. For independent developers selling API access at $0.01 per call, these are acceptable trade-offs. But for an enterprise agent making 10,000 API calls per hour to a SOC 2 compliant SaaS provider, these flaws are fatal.

MPP was specifically built to fill these gaps. The process also starts with the 402 status code. But it is no longer a single signed stablecoin transaction; instead, the client negotiates a session, authorizes a spending limit, and makes repeated calls without settling each request.

The underlying layer of MPP is Tempo, a dedicated blockchain launched on March 18, 2026. Its specifications include: over 100,000 TPS, 0.6 seconds of finality, and no token design (no gas tokens, no native cryptocurrencies). Validation nodes are operated by Stripe-approved partners. This trade-off is evident: Tempo sacrifices decentralization for throughput, cost predictability, and regulatory compliance. Stripe does not hide this.

The partner list at launch clearly indicates its intention to enter the enterprise market. The first batch of over 50 services includes OpenAI, Anthropic, Google Gemini, Dune Analytics, and Browserbase. Design partners include Visa, Mastercard, Deutsche Bank, Shopify, and Revolut. This is an extension of Stripe's existing merchant network into agent business, also equipped with fraud detection (Stripe Radar), tax calculation, and dispute resolution mechanisms.

MPP supports USDC, credit cards, debit cards, buy now pay later, and digital wallets within a single protocol. Agents paying with USDC on the Base network and agents paying with Visa corporate cards will access the same endpoint and receive the same receipt. Most enterprise procurement is still done through bank cards. A payment protocol that mandates only cryptocurrency acceptance is one that the vast majority of CFOs would not approve.

Its weaknesses are also real. Tempo is a new chain with no ecosystem beyond Stripe. Initially, it only accepts cryptocurrency in the U.S. The protocol creates a significant dependency on Stripe's infrastructure. Additionally, MPP's session model introduces state, meaning it faces failure modes that x402's stateless design avoids: session expiration, partial settlement disputes, and authorization token management.

x402 vs MPP

Jumping out of the protocol level, the platform strategy becomes clear. Both companies are building a full-stack agent business ecosystem. The protocol is merely the payment layer in this larger game.

The most critical move in this showdown is Stripe's last action. Stripe supports the use of x402 on the Base network. It hedges its risk by providing this permissionless protocol alongside its enterprise-level protocol. Coinbase has not reciprocated in kind. There is no path for fiat payments in Coinbase's architecture. If an agent needs to pay with a credit card, Coinbase cannot provide a solution.

Stripe's distribution advantage cannot be overstated. Through integration with ChatGPT, MPP reaches 900 million weekly active users (February 2026), including 50 million paid subscribers. When an AI agent built on the OpenAI platform needs to make a purchase, Stripe is the default track. In contrast, Coinbase's distribution relies on developer adoption: the Base ecosystem, member companies of the x402 Foundation, and crypto-native builders who prefer permissionless infrastructure.

However, Stripe's own promotion in the agent business has faced setbacks. The ACP was integrated with Shopify at launch, but initially, only about 12 merchants activated the service. Users could browse products through the agent interface, but when it came to payment, they had to jump to the retailer's website to complete the transaction (according to The Information, March 2026). The agent is responsible for browsing, while humans are responsible for payment. This is not agent business. It is merely a product recommendation engine with added cumbersome steps.

Coinbase's counterattack advantage lies in its neutrality. As a payment processor, Stripe has a commercial incentive to keep transactions on its platform. Coinbase has built x402 as an open protocol and handed governance to the foundation. However, Coinbase's platform narrative is complex: its Commerce product will go offline on March 31, Coinbase Business is hosted and limited to the U.S. and Singapore, and it automatically converts everything to USDC at settlement. This open protocol is pure, but the platform behind it is in transition.

Traditional Giants Strike Back: Visa and Mastercard Breakthrough

As Coinbase and Stripe build from the bottom up, Visa and Mastercard are laying out their strategies from the top down. Both major card networks announced agent payment capabilities in early 2026, revealing what happens when traditional giants with billions of existing user credentials enter an emerging market.

Visa's Intelligent Commerce platform launched with over 100 partners, including Samsung, Perplexity, and Microsoft. Its core innovation is tokenized agent credentials: encapsulating Visa card numbers in agent-specific tokens, accompanied by spending limits, merchant restrictions, and real-time monitoring. Visa processes $14 trillion in funds annually. It does not need to invent a new payment track.

Mastercard's Agent Pay takes a similar approach. Agent tokens provide authorization for delegated spending. Integration with Fiserv introduces thousands of existing merchant acquiring institutions. Santander completed the first agent payment in the EU on March 2, 2026. Mastercard bets that the regulatory and compliance infrastructure it already operates is the hardest nut to crack, and extending it to the agent space is merely incremental engineering development.

Google's AP2 protocol serves as the governance layer above all of this. AP2 has over 60 founding partners, including Adyen, American Express, Ant Group, Coinbase, Etsy, Intuit, JCB, Mastercard, PayPal, Revolut, Salesforce, UnionPay, and Worldpay, defining how agents discover services, exchange authorizations, and delegate permissions. AP2 deliberately maintains payment method agnosticism. An AP2 authorization can approve payments through x402, MPP, Visa, or Mastercard.

What is Happening On-Chain

Data devoid of context is merely a marketing slogan. Here is the background of the situation.

Do not confuse low transaction volume with a lack of prospects. Stripe's real commercial transaction volume in the first year was zero. AWS had negative unit economics in its first five years. Those 75 million x402 transactions are not revenue. They are 75 million proofs that developers believe this protocol is worth integrating before anyone pays them. The vacuum period between building and buying is the norm for infrastructure adoption.

Navigating the Regulatory Gray Area

Every protocol in this stack operates in a regulatory gray area, and this area will not remain gray forever. The EU's MiCA legislation (Markets in Crypto-Assets Act, effective June 2024) imposes licensing requirements on stablecoin issuers and crypto asset service providers. The service provider model of x402 (settling stablecoin payments on behalf of merchants through third parties) fully complies with MiCA's definition of crypto asset service providers. No x402 service provider has yet obtained a MiCA license. This is a ticking time bomb for its adoption in Europe.

MPP faces a completely different regulatory landscape. The EU's PSD2/PSD3 framework regulates payment initiation services. An agent representing a user authorizing a payment session is executing the action of payment initiation. Whether MPP's session model requires a PSD3 license depends on whether the agent is classified as a payment service provider or a technology intermediary. Stripe has a PSD2 license across Europe. The independent MPP implementation does not.

In the U.S., the Consumer Financial Protection Bureau has yet to release specific guidelines regarding AI-mediated consumer transactions, but existing consumer protection rules (E regulations for electronic funds transfers, TILA for credit) apply equally, regardless of whether payments are initiated by humans or agents. The EU payment by Mastercard and Santander (March 2, 2026) was settled under Santander's existing banking license. It did not set a new regulatory precedent. It demonstrated that existing frameworks can accommodate agent payments under the right institutional packaging.

The takeaway for builders regarding regulation is that permissionless protocols will need "licensed packaging" in regulated markets. The open service provider model of x402 works today in the U.S. But without compliance with the MiCA legislation, it will struggle in the EU. MPP's dependency on Stripe is a weakness in the crypto-native circle but an advantage in regulated markets, as Stripe's existing licenses provide compliance assurance. The protocols that win in various jurisdictions will be those that solve regulatory issues, not necessarily those that offer the best developer experience.

What Happens Next

x402 will become the permissionless underlying layer for the internet's long tail market: independent APIs, open data sources, micropayment threshold content, and any use case where creating an account before payment would introduce unacceptable friction. MPP will serve as the session layer for high-frequency, heavily regulated enterprise workloads: SaaS APIs, cloud services, financial data, and anything requiring approval from Fortune 500 procurement teams.

They coexist because they serve different trust boundaries. An anonymous agent scraping market data on a public API does not need Stripe Radar. But a healthcare agent processing insurance claims needs not only Stripe Radar but also compliance with HIPAA regulations, audit trails, and refund logic. Building both functionalities into a single protocol would be too heavy for the first case and too light for the second. A dual-layer design is the correct architecture.

The card networks provide a third layer: scalable compliance and governance. Visa's tokenized credentials and Mastercard's agent tokens will handle situations that require extending existing human financial relationships to their agents. For the vast majority of consumers, their first authorized agent payment will be completed through their existing Visa or Mastercard, not through a stablecoin wallet. This is a simple arithmetic problem: 4 billion bank cards versus 50 million stablecoin wallets.

The End of the Era of Traffic Monetization

For thirty years, the internet's business model has been: show ads to humans, harvest attention, and convert clicks into revenue. Google's $300 billion annual ad revenue, Meta's $135 billion, Amazon's $50 billion. All of this relies on a living person sitting on the other side of the screen.

Agents do not click on ads. They have no attention to harvest. They do not make impulse purchases, do not respond to brand affinity, and do not watch banner videos before accessing API endpoints. They are only responsible for assessing and then paying, or leaving. The advertising model is structurally incompatible with machine users. This is not a prediction. This is an arithmetic fact.

Gartner states that by 2028, $15 trillion of B2B procurement will be completed through AI Agents. Even counting only 10%, that is $1.5 trillion needing payment tracks, settlement, and dispute resolution mechanisms. Bain predicts that by 2030, the scale of the U.S. Agents business will reach $300 to $500 billion. The AI Agents market itself will grow from $7.84 billion to $52.62 billion at a compound annual growth rate of 46.3% (according to Markets & Markets data). This base is large enough for x402, MPP, Visa, and Mastercard to each carve out their own space.

But the following number should make every developer in this field feel uneasy: only 7% of consumers trust third-party AI platforms to shop for them (according to Bain's 2026 survey of 2,016 U.S. respondents). Payment protocols are ready. Infrastructure is ready. But humans are not ready.

Thirty years ago, due to the high costs of micropayments, the internet reluctantly chose the advertising model. Today, the technological barrier for micropayments is nearly zero, but trust has become the most expensive luxury. Whoever can bridge this gap will dominate the next generation of internet commerce. Because payment technology is ultimately just a means; whether users are willing to pay is the real moat.