On September 2, Anthropic released what it calls a blueprint for commerce agents. Think of it as a starter kit: the harnesses, patterns, and guardrails an engineering team needs to get a shopping agent running in days rather than weeks. It ships with working reference implementations for retail, travel, telecom, and ticketing, plus a Claude Code plugin to get you going. The code sits in the open on GitHub.
The timing isn’t accidental. The holiday season is right around the corner, and Anthropic is selling the kit as a shortcut to it. Big names like Shopify, Priceline, and Intuit already build agents on Claude, according to Anthropic. The blueprint is meant to make that repeatable for everyone else.
Two pieces come in the box. The shopping agent lives inside the retailer’s app or website. It searches the catalog, assembles the right items for multi-part requests, remembers preferences, shows products and the cart right in the conversation, and hands off to checkout at the end. Payment stays with the retailer. A customer can say: I need a tent, a sleeping bag, and a stove for a weekend with two kids. The agent takes it from there. It also answers service questions about orders, returns, and refunds instead of pointing to a help page.
The second piece is the merchant agent, built for the people running the store. It answers questions about sales performance, flags looming stockouts on its own, recommends prices and promotions based on the store’s own history, and drafts marketing campaigns. When it proposes a change, a human has to approve it before anything goes live.
Anthropic makes a point of the guardrails: prices and products are tied to actual catalog data, and manipulative upsell patterns are meant to stay out. You can deploy through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. Named partners include Accenture, Mastercard, and Visa.
The numbers Anthropic cites sound good: one partner saw carts grow 30 to 35 percent, with customers roughly 60 percent more likely to finish a purchase. Those figures come from the vendor, not an independent measurement, so take them with a grain of salt. What I find interesting is the approach. Anthropic isn’t selling another model here, it’s selling a recipe. If you already build with Claude, you get reference code that shows how to structure an agent cleanly, down to the details like tool limits and prompt caching that otherwise eat weeks of trial and error. That’s the real lever.
Sources: Anthropic: Building commerce agents with Claude · GitHub: anthropics/commerce-agents · Reuters via Investing.com