Agentic Commerce Splits Into Reasoning and Payment Layers

Anthropic spent the first days of September teaching Claude to build a shopping cart. It spent none of that time teaching Claude to charge a card. That gap is not an oversight. It is the clearest sign yet that agentic commerce is settling into a stack with distinct layers, and that no single company, not even the one building the most capable model, intends to own all of them.

The Blueprint Anthropic Will Not Finish

On September 2, Anthropic released Claude Commerce Agents, an open reference blueprint for building two kinds of agents: a shopping agent that searches a catalog, compares products, and assembles a cart on a customer’s behalf, and a merchant agent that reads sales data, flags inventory problems, and drafts pricing and marketing moves for a store’s staff to approve. The code ships as a public repository with reference implementations across retail, travel, telecom, and entertainment, and it is designed to run on the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI.

Anthropic frames the payoff in commerce metrics, not AI metrics: retailers running shopping agents on Claude have seen carts grow as much as 35% larger, with shoppers 60% more likely to complete a purchase. Priceline is one of the early implementations. “A trip is one of the most complex things a person buys,” the company’s VP of AI Experiences said of the deployment. “Claude models are good at that reasoning.”

What the blueprint does not contain is the interesting part. There is no payment protocol, no checkout flow, and no advertising layer. The shopping agent’s checkout tool renders the finished cart and hands it back to the host application to complete; it never touches a card number or a payment gateway. The merchant agent’s writes are staged, not live, held for a human to approve before any price or listing actually changes. Anthropic built the reasoning layer and stopped deliberately at the edge of money moving.

Two Networks, Two Different Answers to the Same Problem

Visa, Mastercard, Shopify, and Accenture are named as early collaborators on the Anthropic blueprint, but Visa and Mastercard are not waiting for Anthropic to define how an agent actually gets authorized to pay. Both card networks have spent the past year building separate, competing answers to that question, and neither looks like the other.

Visa’s Trust Layer

Visa’s approach, the Trusted Agent Protocol, is built to let a merchant tell a legitimate shopping agent apart from a scraping bot before a transaction ever starts. Mechanically, it works through agent-specific cryptographic signatures built on the HTTP Message Signature standard and aligned with Web Bot Auth, transmitting three elements with each request: an Agent Intent flag showing whether the agent is browsing or actually trying to buy, a Consumer Recognition signal showing whether the shopper is an existing customer, and optional payment information for checkout. Visa has lined up a wide processing and infrastructure bench around the protocol, including Adyen, Ant International, Checkout.com, Coinbase, CyberSource, Elavon, Fiserv, Microsoft, Nuvei, Shopify, Stripe, and Worldpay. “We believe the entire payments ecosystem has a responsibility to ensure sellers can trust AI agents as much as they trust their best customers,” Visa Chief Product & Strategy Officer Jack Forestell said when the protocol launched.

Mastercard’s Machine-Speed Bet

Mastercard’s Agent Pay, and its newer extension Agent Pay for Machines, targets a different failure mode: not fraud at checkout, but a payments system that was never built for the volume and speed autonomous agents can generate. The service is built on four capabilities Mastercard calls credentialing, permissioning, transacting, and settling. Each agent gets its own credentials and a way to prove trusted intent; organizations set spending limits and authorization rules that are enforced programmatically rather than checked after the fact; verified participants transact continuously across providers; and settlement runs across cards, bank accounts, and stablecoins rather than a single rail. More than thirty companies are already integrating, spanning traditional processors like Adyen and Global Payments and crypto-native infrastructure like Coinbase, Polygon, and the Solana Foundation. Mastercard has already run live agent-executed payments with Santander in Europe and expanded the system into Asia-Pacific and Latin America over the first half of 2026.

A Year of Parallel Building, Not Coordination

The arc matters here. Visa put the Trusted Agent Protocol into the market in October 2025, then spent the following two months signing up processors and confirming that hundreds of agent-initiated transactions had cleared successfully, with Visa predicting millions of consumers would use agents to buy something by the 2026 holiday season. Mastercard moved on its own track: Agent Pay launched in April 2025 with Microsoft, IBM, and Braintree as first partners, then spent 2026 rolling out region by region, running Australia’s first fully authenticated agentic transactions in January, completing a live end-to-end agent payment with Santander in Europe in February, and expanding into Latin America and the Caribbean in March, before adding the machine-to-machine layer in June. Neither company waited for the other, and neither waited for Anthropic. The blueprint that landed this week is the reasoning layer arriving roughly a year after the authorization layer had already started shipping.

That sequencing shows in how the two networks talk about risk. Nathan McCauley, chief executive of Anchorage Digital, one of Mastercard’s settlement partners, said the multi-rail approach to Agent Pay for Machines “creates a foundation that our clients can build on with confidence” precisely because it does not force a choice between card rails and stablecoin rails. Farooq Malik, co-founder and chief executive of Rain, another partner, put the same point more bluntly: “The future of payments cannot run through a single closed ecosystem.” Both remarks describe a settlement layer built to hedge against exactly the kind of single-vendor lock-in that Anthropic’s blueprint studiously avoids by staying out of payments altogether.

Why the Stack Split This Way

None of this is coincidence. A large language model is good at the part of shopping that looks like judgment: comparing a two-person tent against a competitor’s, remembering that a customer prefers aisle seats, drafting a markdown campaign for slow-moving inventory. It is a much worse fit for the part of payments that has to be deterministic, auditable, and legally accountable for every dollar that moves. Card networks spent decades building exactly that kind of infrastructure, and neither Visa nor Mastercard shows any interest in handing the authorization and settlement layer to a model provider, agentic or not.

That division is also why the industry has spent 2026 building shared standards around the edges rather than inside any one vendor’s product. EMVCo’s framework for card-based agentic payments, published for industry feedback earlier this month, exists precisely because Visa, Mastercard, and the other card schemes need a common way to certify what an agent is allowed to do before any of their individual protocols can interoperate. It follows the same pattern this publication has tracked as the card networks build shared rules for AI agents rather than compete on incompatible ones, because a merchant that has to integrate five different agent-trust protocols will simply integrate none of them.

What It Means for the Finance Leader

The practical consequence is that there is no single vendor to buy agentic commerce from, at least not yet. A retailer or bank evaluating this space is really evaluating three separate decisions: which model provider builds the reasoning layer, which network or combination of networks handles agent authentication and fraud liability, and which settlement rails, cards, bank transfers, or stablecoins, actually move the money. Anthropic’s blueprint is explicit that the checkout and payment pieces are the deployer’s responsibility, which means a finance team adopting it inherits the job of wiring in Visa’s protocol, Mastercard’s, both, or something else entirely.

Liability is the question worth pressing hardest in procurement conversations right now. Visa’s protocol is fundamentally a trust signal, not a liability shift; Mastercard’s permissioning model enforces spending limits an organization sets itself, which means the organization is still the one that has to set them correctly. Neither company has publicly resolved who eats the cost when an agent, operating exactly as designed, buys the wrong thing at the wrong price. Anthropic’s own design tells on this uncertainty: the merchant agent stages every write for human approval rather than letting it go live automatically, which is a tell that even the company building the reasoning layer does not yet trust the stack enough to remove the human gate on its own side.

The Open Question Nobody Has Answered

The metrics Anthropic and Mastercard both lead with, larger carts, higher conversion, always-on machine transactions, describe upside for the seller. They say nothing about what happens the first time an agent’s spending authorization is compromised, or when a permissioning rule is misconfigured and a machine-to-machine payment loop fires thousands of times before anyone notices. Mastercard’s own pitch for Agent Pay for Machines is explicitly about volume and speed that a human could not review in real time, which is the same property that makes a failure mode expensive before a human can catch it.

For now, the honest read is that agentic commerce has a capable reasoning layer, two well-funded and incompatible answers on trust and authorization, and a settlement layer still being stitched together underneath. A finance leader who waits for one vendor to hand over a finished, single-integration product is going to wait through several more product launches.

The more useful move this quarter is a three-part evaluation, not a vendor shortlist. First, map which of the three layers a given use case actually needs: a customer-facing shopping assistant needs the reasoning layer and a trust protocol; a machine-to-machine procurement workflow needs credentialing and settlement more than it needs conversational polish. Second, pressure-test the liability terms on whichever authorization layer gets chosen before any agent receives a live spending limit, specifically who is responsible when a permissioning rule is set correctly but exploited anyway. Third, treat the human approval gate Anthropic built into its own merchant agent as the current industry consensus on trust, not a temporary training-wheel: every vendor in this stack, including the ones building the model, is still keeping a person in the loop before money moves.

Source: Anthropic