Anthropic has released a reference implementation of commerce-agents on GitHub, under the Apache 2.0 license, which includes one shopping agent and one merchant agent, covering four operational vertical examples: retail, travel, telecommunications, and ticketing, along with the Claude Code plugin. The company states that this is a forkable blueprint and not a maintained product with a service level agreement, and external contributions are not accepted.

The shopping agent is embedded in the merchant's own applications or websites, allowing for natural language search for products, comparison, multi-item combinations, preference tracking, shopping cart assembly, and answering questions about orders and return policies. Checkout is handled by the merchant's own payment page or a third-party payment provider, and the warehouse code does not include direct payment interfaces. The merchant agent monitors sales, inventory, pricing, and marketing, generating pending modifications for price adjustments, restocking, and promotional changes that must be manually approved before being written into production. Prices and products are constrained within a real catalog, with safety, sourcing, pricing adjustments, and approvals intercepted at the tool layer, not just written in prompts. The interface presents tool calls with types rather than pure text.

The architecture opposes breaking down search, returns, and pricing into multiple sub-agents. The official engineering document states that business conversations are tightly coupled cross-intent dialogues, and handoffs between sub-agents would lose shopping carts, pending modifications, preferences, and history, increasing tokens and latency; instead, a single Claude retains the complete dialogue and loads Skills on demand. The company claims that in comparative deployments across multiple enterprises, a single agent with Skills performs better in quality than "a universal prompt" and sub-agent breakdowns, often saving tokens and reducing latency. System prompts account for over 30% of usage, with the remainder going into Skills, while safety, legal, and brand rules always remain in the prompts.

The official blog states that retailers running shopping agents on Claude have seen shopping cart sizes increase by approximately 35%, with a 60% higher likelihood of completing purchases. Product lead Angela Jiang told Reuters that one partner saw shopping cart sizes increase by about 30% to 35%, with a completion probability of around 60%. These figures are based on partner observations and have not been disclosed with sample sizes or independent audits. Notable enterprise clients include Shopify and Priceline. Shopify has also opened a sample library, connecting storefronts and backend interfaces using the same blueprint, with shopping carts still directed to the store's checkout page.

Who is buying and who is selling: The event is driven by making conversations into auditable in-store guidance, not by Anthropic collecting payments. The buyers are retail engineering teams looking to launch guidance before the holiday, while the sellers are using reference code to exchange Messages and Managed Agents calls. Beneficiaries are merchants who still hold checkout and approval rights, while those under pressure are agent shopping products that write payments into model prompts. Funds flow from advertising and customer service conversations to model inference fees, while payment settlements remain with the original card groups and gateways.

Source: Public Information