Anthropic published Claude Commerce Agents on September 3, an Apache-2.0 reference project for retail, travel, telecom, and entertainment companies that want their own shopping or merchant agents.
The kit covers two roles. A shopping agent lives inside a merchant’s own app and manages search, multi-item requests, comparisons, carts, and order questions, with skills for discovery, purchase research, planning, customer care, and personalization. A merchant agent supports store staff with sales insights, inventory alerts, pricing ideas, and campaign drafts.
Anthropic’s core argument is that most teams overcomplicate the build. It warns against intent routers and one-subagent-per-domain designs, noting that every handoff between agents loses state and costs tokens, while a single agent holding the cart, preferences, and history sidesteps both problems. Enterprise deployments showed the single-agent design beating a big-prompt setup and a subagent setup on quality, usually at lower cost and latency.
One set of prompt, skill, and tool definitions feeds every runtime, from the Claude API to the Agent SDK to Claude Managed Agents, and the code also moves to Bedrock, Foundry, or Vertex AI. A Claude Code plugin scaffolds new agents.
Practical guidance fills out the release: components rendered as tools rather than prose, eager streaming against perceived latency, request ordering that reaches 90 to 99 percent prompt caching hit rates, and asynchronous memory extraction that beat in-turn saving by 13 percent on fact recall.
