Research Status Report on AI Payment Protocols: The Emerging Two-Tier Architecture

Within the six months from September 2025 to March 2026, every major player in the global payments industry made major moves. OpenAI and Stripe jointly released the Agentic Commerce Protocol (ACP). Google launched the Universal Commerce Protocol (UCP). Within a week, Visa and Mastercard rolled out their respective Agent payment frameworks in succession. Over the following two months, Coinbase’s x402 protocol processed more than 15 million transactions cumulatively on the Base chain. In March 2026, Stripe and Tempo jointly released the Machine Payments Protocol (MPP).

These dense actions by tech giants and financial institutions are absolutely not a coincidence—they’re a collective response by the payments industry to the same challenge: when AI Agents become the most active consumers on the internet, the existing payment infrastructure can no longer fundamentally meet their operating needs.

Every design of legacy payment systems is built on the premise of “human operation”: relying on browser interfaces, manually filling out forms, clicking “confirm payment,” and verifying identity through CAPTCHAs. Agent operations are fundamentally different: they need machine-readable standardized interfaces, authorization responses at millisecond latency, and infrastructure that supports high-frequency, small-value settlement.

This infrastructure battle won’t be led by a single protocol; instead, a clear two-layer architecture is taking shape. The intent layer defines “who is the merchant and how matching happens,” while the settlement layer defines “how funds actually move.” The two layers are independent, evolve separately, but neither can be missing. Without any one layer, the business closed loop of the Agent economy can’t be achieved.

First Part: Intent Orchestration Layer

The intent orchestration layer converts an Agent’s transaction intent into an end-to-end executable flow: discovering goods or services, adding to cart, and triggering payment. This layer has currently formed two tracks of completely different nature.

1.1 Agents Shopping on Someone’s Behalf

The core contradiction in this track is not payments, but access. Traditional e-commerce platforms are designed for human users; Agents can’t parse visualized pages and can’t click interactive elements. For an Agent to complete purchases on behalf of a human, merchants must provide machine-readable standardized interfaces.

ACP (Agentic Commerce Protocol): AI shopping experience within a closed ecosystem

ACP was jointly released by OpenAI and Stripe in September 2025. Its core mechanism is delegated payments: when confirming a purchase, users grant payment authority to the Agent. The Agent completes the transaction through payment credentials that comply with the Delegated Payment Spec, while the merchant retains the Merchant of Record status. Currently, Stripe’s SPT is the first and only production-ready solution in this system.

ChatGPT Instant Checkout launched in September 2025 but was shut down in March 2026 due to too-low conversion rates. OpenAI has shifted its strategic focus to product discovery—after ChatGPT showcases products, it funnels users to the merchants’ native sites to complete the transaction. The ACP protocol itself is retained in a more streamlined form, supporting exclusive in-ChatGPT apps for several large retailers. Merchants must apply to onboard, and OpenAI controls what is displayed.

UCP (Universal Commerce Protocol): Long-term layout with open standards

UCP was personally announced by Google CEO Sundar Pichai at the NRF Retail Industry Conference in January 2026. It has already secured participation from more than 30 partners, covering mainstream platforms such as Shopify, Stripe, Visa, Mastercard, Walmart, and Wayfair. UCP’s core mechanism is merchants’ capability declarations: merchants publish a JSON-formatted UCP configuration file under their own domain, declaring the transfer and payment capabilities they support, so that AI can read it directly. With this, Google builds Gemini into the core discovery layer for Agent shopping.

The key difference is that Google deliberately avoids the middleman role. It doesn’t need to own the transaction itself; it only needs to control the upstream product discovery process. ACP and UCP are not competing relationships—they represent two different market propositions: the former trades a closed ecosystem for an extremely optimized user experience and more control; the latter trades open standards for a larger scale effect and interoperability.

1.2 Transactions Between Agents

If Track A addresses how Agents can complete shopping on behalf of humans, Track B must solve a deeper underlying issue: when both parties to a transaction are Agents and no human merchants are involved, where does trust come from? Agents lack reputation backstops, and consumer protection regulations can’t be applied. The core contradiction is: how to ensure the reliability of value exchange in a zero-trust environment.

ERC-8183 + ERC-8004: On-chain de-trusting task contracts

ERC-8183 was co-released in March 2026 by the dAI team of the Ethereum Foundation and Virtuals Protocol. Each Job consists of three parties: the Client (delegator), the Provider (service provider), and the Evaluator (assessor). Funds are escrowed by smart contracts until task acceptance is completed. The transaction parties don’t need to trust each other; they only need to trust the contract itself. ERC-8004 is the corresponding identity protocol: once each Agent is registered on-chain, it generates a reputation score based on accumulated historical transactions. Currently, there are about 24,000 Agents registered across the entire network.

Virtuals Protocol’s Butler is the biggest driver of this system. It breaks down composite tasks and assigns them to specialized Agents for execution. This model aims to promote this kind of three-party contract mechanism as an open standard, but large-scale developer adoption still takes time.

The structural differences between the two tracks directly affect the selection of settlement-layer protocols: Track A’s transactions naturally connect to fiat currency channels, while Track B’s transactions naturally rely on stablecoin channels.

Second Part: Settlement Layer

If the intent orchestration layer determines “what is being traded,” then the settlement layer has to solve “how the money arrives reliably.” At present, five protocols are competing in this space, each with different emphases in design thinking and applicable scenarios.

2.1 Delegated Payment / SPT (Stripe)

• Core idea: Extend the existing card payment ecosystem, rather than rebuilding it.

• How it works: After a user authorizes an Agent, Stripe generates a shared payment token (SPT), which is stored by the agent. When a transaction occurs, the agent presents this time-bound token with an amount limit to the merchant. Funds are then settled through Stripe’s standard card payment rails. In the background, Stripe has already connected with Visa’s “AI-Ready” and Mastercard’s “Agentic Token” interfaces. Therefore, regardless of which card organization the underlying transaction routes through, the merchant faces a unified SPT interface.

• Applicable scenarios: Ideal for standard retail and large-value transactions—especially for Agent-to-Agent payments that require consumer protection mechanisms like chargeback.

• Main limitations: Its architecture depends on legacy card networks, so it doesn’t fit high-frequency machine-to-machine payment scenarios involving micro amounts (e.g., below 1 cent).

2.2 Visa Intelligent Commerce and Mastercard Agentic Token

• Core idea: Upgrade tokenization techniques from traditional card networks to adapt to Agent transactions.

• How it works: Replace real card numbers with dynamically encrypted payment tokens. Each token is bound with Agent identity, spending limits, validity periods, and metadata such as the set of merchants it can be traded with. Final settlement still occurs through the existing card network.

• Development status: In September 2025, Mastercard partnered with Australia’s Commonwealth Bank to complete the world’s first fully identity-verified Agent transaction. Visa also completed preliminary deployments in Europe through its “AI Ready” program.

• Main limitations: Minimum fees inherent to card networks are a hard constraint, making it difficult to support future massive micro-payments below $1.

2.3 x402 (Coinbase)

• Core idea: Go back to the internet’s underlying protocol layer and use the “402 Payment Required” status code from HTTP standards—which hasn’t been widely used—to integrate payments natively.

• How it works: When an Agent requests a resource that requires payment, the server returns a 402 response along with payment parameters. After the Agent signs the authorization, the settlement node in the protocol completes an atomic swap on the blockchain (typically using USDC), taking about two seconds. The entire process requires no account registration, API keys, or identity verification.

• Data and current state: As of the end of 2025, this protocol has processed over 100 million transactions across multiple blockchains. But some analysis indicates that a significant portion of the traffic consists of internal protocol testing and looping transactions. Although its architecture can be designed for per-sat/penny-level payments without minimum fee limits, the current challenge lies in increasing adoption density and transaction quality in real commercial use cases.

2.4 Nanopayments (Circle)

• Core idea: As an enhanced solution to x402, it specifically optimizes the economic model for extremely high-frequency, ultra-small payment scenarios.

• How it works: It also triggers the HTTP 402 response, but the settlement layer uses a batching architecture: the payer first deposits USDC into the Circle gateway, and subsequent payments are authorized via off-chain signatures and periodically complete on-chain settlement in batches. This effectively makes gas costs negligible and supports payments as low as one-millionth of a dollar.

• Main limitations: Both transaction parties must pre-open accounts and deposit funds into the Circle gateway, making it, to some extent, a semi-closed system and preventing real-time atomic fund delivery. The protocol began running a testnet in March 2026.

2.5 MPP (Tempo + Stripe)

• Core idea: Build a unified, plug-and-play multi-rail payment framework, while also applying to be positioned as the “official production-ready” solution for HTTP 402.

• Key innovation: It allows developers to integrate multiple payment rails within the same protocol framework, so Agents can choose as needed at the time of transaction:

• Tempo stablecoin rail: supports per-transaction on-chain settlement or off-chain session-based batch settlement.

• Stripe fiat rail: completes card payments via shared payment tokens.

• Direct card-organization rail: directly uses Visa/Mastercard smart tokens.

• Bitcoin Lightning Network: integrated via Lightspark.

• Key feature: MPP introduces the concept of “payment sessions,” similar to OAuth authorization. After an Agent pre-authorizes and tops up within a session, it can execute seamless, continuous real-time payments inside the session without putting every individual transaction on-chain.

• Strategic significance: In this setup, Stripe plays a dual role—it is both a joint drafter of the protocol and one of the payment options provided within the protocol. This means that whether the market ultimately favors an open HTTP 402 system or legacy fiat rails, Stripe can ensure its core payments business is embedded in the future ecosystem.

Third Part: Current Status, Challenges, and Opportunities

3.1 Current Status and Challenges

Over the past six months, all relevant core protocols have gone live, but commercialization has lagged overall. In the settlement layer, x402 leads by transaction volume, but true average daily commercial transaction value is about $28,000; in the orchestration layer, ACP’s core product was shut down due to conversion rates being too low. New protocols such as ERC-8183 and MPP face a common situation where the narrative is ahead of production. This marks a key phase: the protocol architecture is basically complete, but large-scale commercial applications haven’t started yet.

The current core challenge is fragmentation in the intent orchestration layer. Merchants have to deal with multiple independent standards, SDKs, and compliance processes at the same time, driving integration costs high and making expectations unclear. History shows that fragmented markets ultimately produce a unified integration layer, but this time may be different: platforms that control traffic entry points (such as OpenAI, Google, and Microsoft) have strong incentives to build and maintain their own closed ecosystems rather than pushing open integration. This logic is playing out simultaneously across global markets. The end state is likely to be multiple coexisting regional closed ecosystems rather than a single unified open standard. Therefore, in the future, the integration layer will not be led by platforms; instead, it will be built by third-party infrastructure providers that serve merchants.

3.2 Market Opportunities

Based on the above judgment, opportunities are clearly present on two levels:

Settlement layer: the most certain opportunity

No matter how fragmented the upper ecosystems are, payments are the underlying issue that any Agent must solve. One clear trend is that the orchestration layer continues to fragment due to platform interests, while the settlement layer is moving toward abstraction and integration due to developers’ efficiency pressures. Developers can’t maintain separate payment integrations for every ecosystem, and the economic motivation to consolidate into a unified solution keeps strengthening.

This creates clear requirements for Agent wallets: they must support multiple payment rails. Fiat rails (such as SPT and Agentic Token) cover traditional real-world consumption; stablecoin rails (such as x402 and MPP Session) cover on-chain services and A2A transactions. The two coexist and, in the short term, won’t merge. The responsibility for flexible adaptation lies on the Agent side, not the merchant side: merchants choose which payment rails to support—that’s their relatively stable decision. Enterprises only need to configure stablecoins and authorize cards for Agents, and then Agents can complete payments according to the rails supported by counterparties. Only a wallet that can handle multiple rails can cover the full range of Agent consumption scenarios. Its value will accumulate continuously as each transaction crosses different ecosystems, forming a deep infrastructure moat.

A2A economy and business model reconstruction: a long-term blue ocean direction

The real market gap is in the application layer services. Today, the A2A economy is still limited to crypto-native scenarios. Technically, having an Agent “hire” another Agent to complete real-world tasks (such as data analysis, content creation, legal research, and code review) is fully feasible, but the supply of corresponding on-demand, per-call API services is extremely scarce. This is the biggest long-term opportunity—and also the area with the fewest competitors right now.

This opportunity is currently constrained by a real cold-start problem: reputation-based trust mechanisms like those in ERC-8183 need enough transaction density to generate meaningful trust signals. Microsoft predicts that by 2028, the number of active AI Agents will reach 1.3 billion, but the current number is orders of magnitude away from that goal. This isn’t a temporary gap that will naturally dissolve; it’s a threshold A2A economics must cross to expand beyond crypto.

Its deeper significance is business model reconstruction. The mainstream internet advertising and subscription models are built on the assumption that “users are people.” Agents don’t care about ads and don’t need monthly subscriptions; they pay only for the results of single tasks. The “pay-per-call” model represented by HTTP 402 provides API service providers with a new path: shifting from selling access permissions to selling exact results, enabling more granular value exchange. The expansion of the A2A economy and the adoption of HTTP 402 are essentially two sides of the same proposition.

Conclusion

Agent commerce will evolve along two dimensions in parallel: on the consumer side (Agents shopping on behalf of users), it will mainly rely on card rails, with development depending on establishing enterprise authorization and user trust. Between Agents (A2A) is technically ready on the stablecoin rails, waiting for the application and services layer to scale up.

The final structure will be a coordinated evolution of a two-layer protocol stack: the intent orchestration layer determines how transactions happen, and the settlement layer ensures how value flows.

For builders, the key right now is to construct broad integration and access capabilities. Infrastructure that can automatically route cross-protocol transactions and shield developers from underlying complexity will gain structural advantages when the market explodes. This is a kind of value that accumulates silently and is hard to replace.

The key to the tipping point will be the moment enterprises are willing to delegate spending authorization to Agents—including auditable transaction traceability, budget delegation mechanisms, and clear responsibility allocation when Agents make mistaken purchases. At that time, Agent wallets that support multiple payment rails and easy-to-use “pay-per-call” service catalogs will become the two most critical pieces of infrastructure that are still unclaimed. Neither of these positions currently has a strong incumbent, and both will become crucial at the same time.