Anthropic Released Claude Commerce Agents: An Apache-2.0 Blueprint for Shopping and Merchant Agents Across Retail, Travel, Telecom and Entertainment

Anthropic has published a commerce blueprint in code, not just in slides. The new repository, anthropics/commerce-agents, ships with two agents — one for shoppers and one for merchants — plus four runnable versions aimed at retail, travel, telecom and entertainment.

This is Apache 2.0, runs locally on Python 3.11+ and Node 22 with an ANTHROPIC_API_KEY, and the same code can point at the Claude API, Amazon Bedrock, Microsoft Foundry or Google Cloud Vertex AI. Anthropic also wrapped the project in two write-ups: a product announcement and a deeper engineering guide on how its commerce agents are put together.

The shopping agent is built for a merchant’s own app. It can search a catalog, handle multi-item requests, compare options, build a cart, and answer questions about orders and returns in the same thread. Anthropic gives it five skills: search-discovery, purchase-research, planning-goals, customer-care and memory-personalization. The merchant agent is for staff instead of shoppers, handling sales questions, inventory alerts, pricing and promotion ideas, and campaign drafts through its own set of skills.

The interesting part is what Anthropic says not to do. It argues against routing every intent to a different subagent, and against splitting a shopping session into separate conversations that lose the cart, preferences and history along the way. In its deployments, one agent with skills reportedly beat both a giant prompt and a subagent setup on quality, often with lower cost and latency. Subagents still have a role, but mostly for narrow jobs like deep research.

Anthropic also treats UI as tooling, not text. Instead of making the model spit out custom tags, the blueprint turns pieces like product cards, itinerary blocks and comparison views into typed tools the server can validate before the client renders them. That keeps the conversation history native, and it makes it easier for the agent to refer back to things like “the first hotel” without extra parsing. The company leans on prompt caching, too, saying the best setups hit 90–99% cache rates when volatile details are kept out of the top of the prompt.