Anthropic's Claude Commerce Agents: What They Mean for Merchants

What Did Anthropic Launch With Building Commerce Agents With Claude?

On September 2, 2026, Anthropic released Building commerce agents with Claude, an open-source blueprint for constructing AI shopping and merchant agents rather than a hosted product you switch on. It ships as Apache 2.0 licensed code on GitHub, containing the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days. Anthropic

The release packages several things a team would otherwise build from scratch:

• Two reference agents: complete, working implementations of a shopping agent and a merchant agent, both grounded in a business's own systems rather than in generic web data.
• Four example verticals: retail, travel, telecom, and ticketing, each with domain-specific flows so a team can start from the closest match instead of a blank file.
• A Claude Code plugin: a commerce-builder plugin that scaffolds a new agent, adds flows, and authors evaluations from natural-language prompts.
• Deployment flexibility: the code runs where teams already build with Claude, including the Claude API, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI.
• Guardrails in the harness: fencing, provenance gates, caps, memory validation, and a merchant approval gate that run inside the tool call, not as prompt instructions.

Anthropic's framing is that a team can see the agents running in a self-guided demo before writing any code, then use Claude Code to customize them to their own catalogs, policies, and brand. The emphasis on days rather than months is the core claim, and several launch partners independently reported setup times measured in minutes to hours. Anthropic

The strategic read is that Anthropic is not selling a storefront or a checkout button. It is publishing the agent brain and the integration scaffolding, then leaving the catalog, the payment rail, and the brand experience to the merchant, which positions Claude as the reasoning layer beneath many different commerce experiences rather than a destination that sits between a brand and its customer. For an ecosystem that has spent the past year debating checkout and payment standards, releasing the agent layer as free reference code is a bid to make Claude the default reasoning engine beneath all of them. Anthropic

What Can a Claude Shopping Agent and Merchant Agent Actually Do?

The blueprint splits commerce into two roles: a customer-facing shopping agent that lives inside a merchant's app or website, and a business-facing merchant agent that supports the people running the store. The shopping agent covers catalog search, comparison, cart building, and order questions, while the merchant agent analyzes performance and acts on inventory, pricing, and marketing. Anthropic

The shopping agent is deliberately scoped to stop short of the money. Anthropic's blueprint provides integration points for catalog, cart, checkout, customer preferences, and order history, and leaves payment to the merchant, whether that is an existing checkout or an agentic payments provider. The table below maps what each agent handles.

Anthropic's own examples make the split concrete. A customer can say they need a tent, sleeping bag, and stove for a weekend trip with two kids and the shopping agent assembles the set, while a store operator can ask what to discount to clear last season's inventory. The shopping agent's guardrails constrain prices and products to real catalog data and avoid manipulative upsell patterns, and the merchant agent proactively flags problems such as an item about to sell out before a promotion starts. Anthropic

Dimension	Shopping agent	Merchant agent
Primary user	Customers shopping in the app	Staff running the store
Core job	Search, compare, plan, build a cart	Analyze sales and act on operations
Example task	Plan and assemble a multi-item order	Recommend what to discount to clear inventory
Data it reads	Catalog, cart, preferences, order history	Analytics, catalog, inventory, pricing, campaigns
Write authority	None; hands the cart to checkout	Staged; every change waits for human approval
Payment	Left to merchant checkout or a payments provider	Not applicable

The division matters because it isolates risk. A customer-facing agent that cannot move money and a back-office agent that cannot publish a change without sign-off are both safe to deploy quickly, which is a large part of why partners report standing agents up in hours rather than months. Anthropic

How Are Claude Commerce Agents Actually Built?

Anthropic's companion engineering guide argues for one model in a standard agent loop, with skills for the long tail and tools that call the systems you already run, rather than a network of orchestrated subagents. The design keeps latency and cost predictable while still covering a wide surface of commerce tasks. Anthropic

Six patterns define the reference architecture:

• Skills over subagents: per-domain modularity comes from Agent Skills such as search, purchase research, and customer care, which give separation of concerns without the handoff tax of multi-agent routing.
• Presentation tools as UI components: each surface, from a product grid to an itinerary, is a typed tool the model calls, so the agent renders real interface elements in the conversation instead of loose text. The same tools give the agent a record of what is on screen, so when a customer says the first hotel or the third one down, the layout is already in the conversation.
• Safety in code, not prompts: no model tool call moves money on its own, and order placement and merchant writes end in an action the harness controls through maker-checker flows.
• MCP for backend access: the agents reach catalog, cart, analytics, and finance systems through the Model Context Protocol, with optional connectors to systems like Snowflake, BigQuery, Stripe, and Square.
• Snapshot-based evals: tests assert against captured agent state rather than replaying whole conversations, which makes agent behavior measurable before it ships.
• Memory that survives the session: long-term memory lives in the systems you operate rather than in the model, so a shopper who states a constraint in one visit does not repeat it in the next, keyed per operator for merchant agents and switchable off per deployment where regulation demands it.

Two performance choices are worth copying even outside commerce. Anthropic structures each request into global, session, and volatile segments to reach prompt-cache hit rates in the 90 to 99 percent range, and it extracts long-term memory on a separate asynchronous thread so personalization never adds latency to the customer's turn. Anthropic

Teams choose one of three runtimes for the same agent logic, which is what lets the blueprint scale from a laptop demo to a managed deployment without a rewrite.

Runtime	What it is	Best for
Messages API	A turn loop built directly on Claude's core API	Full control over the loop and existing Claude API stacks
Agent SDK	Anthropic's agent framework with console interfaces	Faster setup with eager tool dispatch and built-in scaffolding
Claude Managed Agents (beta)	A hosted harness that runs the agent and calls MCP servers	Teams that want Anthropic to run the loop, sandbox, and state

All three paths share the same guardrails: fencing, provenance gates, caps, memory validation, and the merchant approval gate run inside the tool call and hold whichever runtime a team picks. Anthropic, Anthropic

Who Is Building on It, and Why Is Trust the Center of the Pitch?

Anthropic launched the blueprint with a roster of commerce, payments, and platform partners, and framed trust as the precondition for agentic commerce rather than a feature of it. The named partners are Shopify, Priceline, Accenture, Mastercard, Visa, Intuit, Klaviyo, Wix, Zomato, Fetch, and Square. Anthropic

The partner statements cluster around several themes: trust, speed, category fit, and consumer demand.

• Payment networks lead on trust: Visa's Jack Forestell said AI will fundamentally reshape commerce but trust must remain at the center of every transaction, while Mastercard's Sherri Haymond called trust the currency of commerce and even more critical in the agentic era.
• Platforms lead on speed: Wix's Dror Zalika said its engineers had a working commerce agent taking prompts within fifteen minutes, and both Zomato and Fetch reported the blueprint running with no blockers, turning days of scaffolding into hours.
• Operations teams see time returned: Square framed its agents as watching sales, labor, and inventory and returning real next steps while keeping sellers in control, and Klaviyo described Claude turning customer data into campaigns and personalization its teams could not action by hand.
• Complex categories show the clearest fit: Priceline's Cobus Kok said a trip is one of the most complex things a person buys, with flights, hotels, and cars to weigh at once, and that the latest version of its Penny assistant runs on Claude because that reasoning is what the models are good at.
• Financial workflows join in: Intuit's Chris Kasten described building a financial system of intelligence that pairs frontier AI reasoning, including Claude, with Intuit's proprietary data so customers can see what is shifting in their business and why.
• Consumer demand is the backdrop: Accenture's Kath Gramling cited research that 85% of consumers are now open to collaboration with an AI agent and nearly three in four would trust a personal AI agent more than their best friend to make a purchase.

The trust framing is not incidental. Anchoring the launch with the two largest card networks signals that agentic checkout will run on the same identity, security, and dispute infrastructure as today's payments, and that merchants want control over how an agent engages their customers before they hand it the cart. Anthropic

What Does This Mean for Merchants and Agentic Commerce Readiness?

The launch turns a soft trend into a concrete bar: businesses now compete to be the store an agent can actually read, search, and transact against, not just the brand a human can find. Anthropic reports that retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase. Anthropic

That demand is not hypothetical. Accenture's Talk to my AI agent study, which surveyed 25,590 consumers across 16 countries, found that nearly three in four (74%) would trust a personal AI agent more than their best friend to make a purchase on their behalf. Accenture

For a merchant, the practical implications sort into five:

• Transactable beats discoverable: being mentioned in an AI answer is worthless if the agent cannot check your inventory or build a valid cart, so machine-readable catalog and availability data becomes a revenue prerequisite, not a nicety.
• MCP is the new integration surface: exposing catalog, cart, and order systems through the Model Context Protocol is what lets any Claude-based agent, yours or a partner's, act on your data.
• Guardrails are a commercial feature: price and product fencing, caps, and approval gates are what make a brand willing to let an agent represent it, which is why Anthropic builds them into the harness.
• Owned agents change the funnel: a shopping agent embedded in your own app captures high-intent conversations on a surface you control, rather than sending the customer into a third-party assistant.
• Measurement has to change too: agent-referred visits and revenue do not appear in a conventional funnel, so tracking how often assistants cite, recommend, and transact against your brand becomes the scoreboard that tells you whether the readiness work is paying off.

The deeper shift is where discovery happens. As consumers delegate research and comparison to assistants, the product page a brand spent a decade optimizing for human eyes is increasingly read by an agent instead, and the agent judges the brand on whether it can retrieve a real price, a real delivery date, and a valid configuration. A brand that reads well to humans but returns nothing usable to an agent simply drops out of the recommendation. The same logic applies inside a merchant's own assistant, which is only as good as the systems it can query, so the readiness work is what turns a shopping agent from a demo into a channel. Anthropic

This is where the readiness work maps cleanly onto an existing model. The Network and Convert phases of the A.G.E.N.T.I.C. Framework describe exactly this progression: implement the protocols that let agents reach your systems, then optimize for transactions an agent can complete end to end. Anthropic's blueprint is the supply side of that equation, and the merchant's data readiness is the demand side.

None of this means starting from scratch. Anthropic ships example implementations for retail, travel, telecom, and ticketing plus a Claude Code plugin that scaffolds an agent from a prompt, so a team adapts reference code rather than building the loop by hand. The harness, the guardrails, and the evaluation patterns already exist, and the readiness gap that remains is almost entirely on the merchant's side of the integration. Anthropic

Where Does This Sit in the Agentic Commerce Protocol Stack?

Anthropic's blueprint occupies the agent-logic and integration layer, deliberately leaving checkout and payment to separate protocols, which is why it names the Model Context Protocol for backend access and hands payment to the merchant. Understanding the stack prevents the common error of treating every agentic commerce standard as interchangeable. Anthropic, Model Context Protocol

Three layers sit beneath a working commerce agent, and Anthropic's release addresses only the first directly:

• Integration and reasoning (MCP): the Model Context Protocol, Anthropic's open standard, is how the agent reaches catalog, cart, and analytics systems, and it is the layer the commerce-agents blueprint is built on.
• Checkout (ACP and UCP): the Agentic Commerce Protocol, co-developed by Stripe and OpenAI, and the Universal Commerce Protocol, led by Google with Shopify and other retailers, standardize how an agent places an order across merchants.
• Payment authorization (AP2 and card-network rails): Google's Agent Payments Protocol and the agentic programs from Visa and Mastercard evidence that a human authorized a specific purchase, which is the trust layer the launch partners emphasize.

The takeaway is that Anthropic is not competing with the checkout and payment protocols, it is sitting above them. A merchant still needs a checkout path and a payment rail, and the blueprint is explicit that payment is left to an existing checkout or an agentic payments provider. The winning stack combines a capable agent, MCP-exposed systems, and a trusted payment protocol, not any one of them alone. Stripe, Google, Google Cloud