Anthropic open-sources Claude Commerce Agents: two agents, four runnable verticals
The Apache-2.0 release ships a customer-facing shopping agent and a staff-facing merchant agent, plus four runnable verticals and a Claude Code plugin. Nothing in it places an order, charges a card or changes a live listing.
Anthropic has open-sourced Claude Commerce Agents, a blueprint for teams that want a shopping assistant in front of customers and a back-office assistant behind the counter. The developer account @ClaudeDevs announced it on 2 September 2026, with runnable demos for retail, travel, telecom and entertainment. The shopping agent sits inside a company's own app: it searches, compares, plans, fills a cart, answers order and policy questions and remembers what the customer said. The merchant agent is the staff side, covering stock, pricing and campaigns. Each is defined once and runs on the Messages API, the Claude Agent SDK or Managed Agents.
One model, one loop
The accompanying guide argues against spreading the work over many sub-agents. Anthropic's recommendation is a single model in a standard agent loop, extended with skills and tools; by its own quality measurements this beats fragmented alternatives while lowering latency and cost. Instructions needed constantly stay in the system prompt, the rest loads as skills. Prompt caching changes the arithmetic: Anthropic reports hit rates of 90 to 99 percent in commerce deployments, and a cached read costs a tenth of a fresh token. The company also passes on numbers from its customers: carts up to 35 percent larger and shoppers 60 percent more likely to complete a purchase, at retailers running shopping agents on Claude. These are figures Anthropic reports rather than independent measurement.
Safety gates and fictional data
The Apache-2.0 repository, in Python, is deliberately inert. Every company, brand, product and person in it is fictional, checkout only hands a finished cart to the host application, and every merchant-side write waits for a human. The guide carries that caution into production: financial actions are staged and run only after human or policy approval, writes accept only server-generated identifiers, and third-party text such as reviews and seller messages is cleaned before it reaches the model. Agents are evaluated with conversation snapshots rather than full simulated dialogues, refusals and prompt-injection scenarios included. A Claude Code plugin scaffolds an agent against an existing backend and reviews one that already runs. Disclosure: GlobalFeed's own engine is Claude.
