Claude’s commerce blueprint ships two reference agents

Anthropic said on September 2 that retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase. Alongside those figures it published a blueprint that any engineering team can fork to build the same kind of agent.

The announcement on the Claude blog describes the blueprint as the harnesses, patterns and guardrails needed to get a commerce agent running in days. A harness is the code around the model that runs the loop, calls tools and enforces limits. The repository has two reference implementations, a shopping agent and a merchant agent, for retail, travel, telecom and ticketing platforms, and a Claude Code plugin to start from. Both headline percentages are Anthropic’s own reported figures and not an independent benchmark, PYMNTS pointed out.

The package builds on what Anthropic learned by watching retailers build and run custom agents on its platform, Digital Commerce 360 reported, and turns that into pre-built designs for agents that serve specific roles in e-commerce. Shopify and Priceline are among the customers Anthropic names as already running agents on Claude that let consumers search in plain language, compare and buy.

What the shopping agent does

The shopping agent sits inside the retailer’s own app or website. The blueprint wires it to five integration points: catalog, cart, checkout, customer preferences and order history. Anthropic’s example request is “I need a tent, sleeping bag, and stove for a weekend trip with two kids.” The agent searches the catalog, assembles a set of items that covers the whole request, adjusts its suggestions to the preferences it remembers, builds the cart and hands it to checkout.

Products, comparisons and the cart appear inside the conversation as interface elements, not just as text. Customer service runs in the same thread: where an order is, how to return or exchange an item, what the refund policy says, instead of sending the customer to a support page.

In the repository these behaviours ship as skills and tools for catalog search, multi-item planning, deep research, personalization, customer care and in-conversation UI. Guardrails are designed to constrain prices and products to actual catalog data and to avoid manipulative upsell patterns.

A person signs off

The merchant agent works for the people running the store. Anthropic’s sample question is “what should we discount to clear last season’s inventory?”, answered from the store’s own data. The agent reports what is selling and what isn’t, tracks inventory, and flags problems proactively, such as an item about to sell out before a promotion starts. It recommends pricing and promotions based on the store’s sales history and drafts marketing campaigns for the products that need to move. The repository packages this as skills for sales analytics, catalog and inventory management, marketing and promotions, plus in-portal UI such as charts and dashboards.

When the merchant agent proactively suggests a change, a person approves it before anything goes live. A newsletter by Linas lists the release’s safety controls as server-issued IDs, transaction caps, staged changes and approval gates, and says the release also includes eval patterns and four runnable vertical examples.

Square is building similar tools. Willem Avé, Square’s head of product, describes “agentic tools that watch sales, labor, and inventory and come back with real next steps, not just an answer, while keeping sellers in control,” and adds: “Trust is the hardest part of that work.”

Fifteen minutes to a first prompt

Teams can build the reference agents on the Messages API, the Agent SDK or Claude Managed Agents, which is still in beta. The code deploys wherever a team already runs Claude: the Claude API, Amazon Bedrock, Microsoft Foundry or Google Cloud Vertex AI. Each vertical has a self-guided demo that runs before anyone writes code. Customization to a store’s catalog, policies and brand then happens through Claude Code.

Early testers mostly reported on setup time. Dror Zalika, head of commerce at Wix, says the company’s engineers “had a working commerce agent taking prompts within fifteen minutes.” Akhil Bansal, a senior engineering manager at Zomato, says the setup “worked exactly as documented.” He also recognized its practices, “from tool iteration limits to prompt caching,” from building Zomato’s own agent.

Ashley Nader, a staff product manager, says her team had both agents running locally “in well under an hour, with live conversations working on the first attempt.” The Claude Code workflow did not give the same output twice, though: “We ran the Claude Code workflow twice and got two different architectures back, each designed to what we’d asked for.”

The repository is at github.com/anthropics/commerce-agents. An engineering deep-dive on how it was built sits at claude.com/blog/the-anatomy-of-effective-commerce-agents.

The cart is where it stops

Payment is left to the retailer, through its existing checkout or an agentic payments provider. PYMNTS goes further: the blueprint carries no payment protocol, no checkout, no ad layer and no product catalog. Those stay with retailers and with partners such as Visa and Mastercard, so no transaction passes through Anthropic. In Shopify’s implementation, the outlet adds, the agent finds products, builds a cart and then passes the shopper to the store’s existing checkout. Linas’s newsletter makes the same point at the code level: nothing in the reference implementation places an order or charges a card.

Shopify has published a public code repository that shows how the agents connect to a live storefront. Vanessa Lee, VP Product, describes “a reference storefront implementation that connects them to a merchant’s store through Catalog, UCP and Shop Sign-in,” with which merchants can build agents that “help customers find products, check out, and answer questions about their orders.”

Competitors are already inside the transaction. Per Linas, Google and Shopify run live agentic checkout through the Universal Commerce Protocol, and Microsoft’s Copilot Checkout counts more than 500,000 participating merchants. The newsletter also says OpenAI retired ChatGPT’s in-chat Instant Checkout in March 2026, roughly five months after launch, after Walmart found that in-chatbot conversion ran about three times below its own site. OpenAI’s payment layer for agent purchases, spotted in ChatGPT’s code, had not shipped as of August 14.

Accenture, Mastercard and Visa are the partners Anthropic names for helping clients and merchant communities adopt the blueprints. Jack Forestell, Visa’s chief product and strategy officer, said: “Merchants are telling us they want more control over how AI engages their customers.”