The Protocol War Nobody Wins

There are three protocols for how an AI agent buys something on behalf of a human, and they don't agree on much. The two biggest payment processors in the world chose opposite strategies for dealing with them.

Stripe co-wrote one of the protocols. Adyen built a layer that translates between all three. The obvious framing is that they're competing, but the actual split is more structural than competitive, and it tells you something about where each company thinks value accrues in agent commerce.

Three protocols, three layers

Before comparing what Adyen and Stripe built, you need to understand what they're building on top of. There are three protocols that matter, and they don't compete as cleanly as headlines suggest.

UCP (Universal Commerce Protocol) is Google's. It launched in January 2026 at NRF with Shopify, Walmart, Target, Wayfair, and Etsy as co-developers. More than 20 companies endorsed it, including Visa, Mastercard, Adyen, Stripe, American Express, Best Buy, and The Home Depot. Published on GitHub under Apache 2.0.

UCP covers the full commerce lifecycle: discovery, cart, checkout, order management, and post-purchase. Merchants advertise their capabilities at a well-known path, and the protocol is designed so that the payment processor is a pluggable component rather than a fixed dependency.

ACP (Agentic Commerce Protocol) is Stripe and OpenAI's. It shipped in September 2025 under Apache 2.0 and originally powered Instant Checkout in ChatGPT. Where UCP spans the full lifecycle, ACP focuses on the checkout session. An agent gathers the buyer's intent, starts a checkout with the merchant, and completes it using a Shared Payment Token.

The SPT is the mechanism that makes ACP work. Each token is scoped to one seller, one currency, one maximum amount, and one expiration. The agent never sees raw card numbers. The merchant charges it through a standard Stripe PaymentIntent, and Stripe enforces the token constraints in real time.

AP2 (Agent Payment Protocol) is Google-originated, now FIDO-governed. Google donated it in April 2026, and Mastercard contributed its Verifiable Intent framework at the same time. AP2 handles authorization through mandates: cryptographically signed credentials that record what a user has approved an agent to do.

Mandates come in three types (Intent, Cart, and Payment), and version 0.2 added support for autonomous "human not present" payments based on pre-authorized mandates.

UCP handles discovery through fulfillment, ACP handles checkout, and AP2 handles authorization, but they overlap where the agent presents payment and the merchant confirms the order.

Stripe's bet: own the checkout

Stripe's position is straightforward: they co-wrote ACP with OpenAI, and the protocol routes checkout through Stripe's infrastructure via the Shared Payment Token. When an agent completes a purchase through ACP, it submits the SPT, the merchant creates a Stripe PaymentIntent, and Stripe processes the charge.

This is the same playbook Stripe has run since 2011: make the integration so easy that developers choose Stripe by default, then own the transaction. ACP is open-sourced under Apache 2.0, but the reference implementation, the SPT, and the checkout API all belong to Stripe.

The bet is that protocol-level integration creates stickiness. Once ChatGPT and other agents use ACP for checkout, merchants who want agent traffic need Stripe. The protocol is technically open, other payment handlers could implement it, but the gravitational pull toward the reference implementation is strong.

Stripe also supports UCP. Their docs include a Stripe Payments handler for UCP, so a merchant using UCP for discovery can still route payment through Stripe. They're positioning to be the payment layer underneath whichever protocol wins.

Adyen's bet: own the translation

Adyen's position is different. They launched Adyen Agentic on June 16, 2026, positioning it as a universal translator for agent commerce.

The platform has three API layers. Agentic Feed exposes a machine-readable product catalog and inventory. Agentic Cart handles checkout orchestration, including tax, fulfillment, and order management. Agentic Payments covers authentication, tokenization, fraud controls, and merchant-of-record flows.

The key design decision is that these layers work across UCP, ACP, and AP2. A merchant integrates once with Adyen Agentic, and their products are available to agents regardless of which protocol the agent speaks. If a customer's agent arrives through Google's AI shopping and uses UCP, Adyen handles it. If it arrives through ChatGPT and uses ACP, same integration.

At launch, it's available only to enterprise merchants in the United States. Meta's AI checkout is the only platform named as compatible, with more to follow.

Karan Katyal, Adyen's Global Head of Agentic Commerce, rates current agentic commerce maturity at 0.5 on a five-point scale. He projects about 1.5 a year from now. Volume in 2026 is effectively zero. But fragmentation will continue for quite some time, and that's the condition Adyen is building for.

Adyen's bet is that protocol fragmentation persists. If merchants need to support multiple protocols, they need a translation layer. Adyen wants to be that layer the same way it became the translation layer for payment methods and currencies.

Where they actually overlap

Both Stripe and Adyen want to process the payment. The protocols are new plumbing, but the revenue model is the same: take a cut of the transaction.

Think of it like two approaches to running a restaurant that serves food from multiple delivery apps. Stripe wrote the menu format for one of the apps and processes the payment when orders come through it. Adyen built a kitchen system that takes orders from all the apps and routes them to the same stoves regardless of which app the customer used.

The difference is where each company positioned itself in the stack.

Stripe builds from the protocol up. Co-authoring ACP gives them influence over how agents present payment. The SPT is a Stripe object. The checkout flow runs through Stripe's API. Even when Stripe supports UCP, it's as a payment handler underneath someone else's protocol.

Adyen starts from the merchant side. The three-layer API abstracts away protocol differences. The merchant's catalog, checkout flow, and payment processing all go through Adyen. The protocol the agent uses becomes an implementation detail that Adyen handles.

I covered Adyen's H1 earlier. Their strategy has been consistent: sit behind the merchant, abstract complexity, charge for the abstraction. Agentic is the same logic applied to a new surface.

The overlap shows up in checkout, where both companies want to process the transaction. If a merchant uses Adyen Agentic and an agent arrives through ACP, does the payment route through Stripe (because ACP's SPT is a Stripe primitive) or through Adyen (because the merchant's payment processing runs on Adyen)?

That question doesn't have a public answer yet. The protocols are too new and the transaction volume is too low for the edge cases to have surfaced. But it's the architectural tension that will define how agent commerce payment processing shakes out.

The trust problem underneath

The protocol competition assumes agents will buy things, but the data suggests that's still an assumption.

Product.ai's October 2026 Trust in AI Commerce Report found that about 10% of US online shoppers who use an AI assistant act on its recommendation without checking another source first. The other 90% verify before buying.

That verification step breaks the agent commerce model. An agent that gathers intent, finds a product, and completes checkout only works if the human trusts it to buy the right thing. When 90% of shoppers double-check, the agent doesn't complete a purchase autonomously. It compiles a shortlist that the human reviews.

This is consistent with what I covered in the card networks piece: the payment infrastructure is being built, but the authorization model, who approves what the agent does, is still unresolved. AP2's mandates are one answer: the human pre-authorizes specific constraints, and the agent operates within them. But mandates require the human to define constraints before the shopping begins, which assumes a level of intent clarity that browsing doesn't have.

Katyal's 0.5-out-of-5 maturity rating reflects this. He said the AI was never the hard part. The hard part is the plumbing underneath: machine-readable catalogs, trust, fraud, liability, and protocol reconciliation.

What this means for merchants

If you're a merchant deciding how to approach agent commerce, the protocol question isn't actually the first question. The first question is whether your product catalog is machine-readable.

Agents can't buy what they can't parse. UCP, ACP, and AP2 all assume structured product data: prices, availability, variants, shipping options, return policies. A merchant whose catalog lives in unstructured HTML pages isn't ready for any of the protocols.

After the catalog, the decision tree branches. If most of your agent traffic will come through Google's AI surfaces, UCP is the starting point. If it comes through ChatGPT, ACP matters. If you're large enough to expect traffic from both, Adyen's translate-everything approach makes more sense than building two separate integrations.

But Katyal's timeline suggests there's no urgency. Volume is near zero, the protocols are still evolving, and nobody has resolved how UCP and ACP's checkout overlap settles while AP2's mandates remain in FIDO's standards process.

Starting with structured product data and waiting for the protocol picture to clarify is probably the right call. Betting everything on one protocol probably isn't.