Anthropic releases Claude Commerce Agents, open-source blueprint for shopping assistants

Anthropic has launched Claude Commerce Agents, an Apache 2.0 open-source reference blueprint designed to simplify the development of AI shopping and merchant assistants. The release includes pre-built agents for retail, travel, telecom, and entertainment. Rather than relying on complex multi-agent routing, Anthropic introduces a single-agent architecture with modular 'skills' to preserve conversation state and reduce latency. The blueprint also optimizes costs via structured prompt caching and treats UI components as typed tools.

Claude Commerce Agents release

• ▪The Claude Commerce Agents blueprint includes four runnable industry verticals: retail, travel, telecom, and entertainment.
• ▪Anthropic released a Claude Code plugin called commerce-builder to scaffold or review commerce agents using command-line tools.
• ▪Anthropic released an Apache 2.0 open-source reference blueprint called Claude Commerce Agents, which contains a shopping agent and a merchant agent.
• ▪The merchant agent in the Claude Commerce Agents blueprint supports store staff with sales performance insights, inventory alerts, pricing recommendations, and marketing campaign drafts.
• ▪The shopping agent in the Claude Commerce Agents blueprint handles catalog searches, multi-item requests, cart building, and customer care within a merchant's application.

Skills versus subagents architecture

• ▪Anthropic argues against using multiple subagents or intent routers in commerce applications, stating that handoffs between subagents cause state loss, higher token costs, and increased latency.
• ▪Anthropic's agent skills architecture loads specific skill instructions into a single agent holding the conversation history, which outperformed subagent and single-prompt designs in enterprise deployments.
• ▪In the Claude Commerce Agents architecture, approximately one-third or more of traffic goes into the system prompt for safety and brand rules, while the rest is handled by dynamic skills.

UI components as tools

• ▪The Claude Commerce Agents blueprint supports an eager input streaming configuration that executes tool calls as arguments stream, reducing perceived latency to a few hundred milliseconds.
• ▪Modeling user interface components as tools allows the Claude Commerce Agents to resolve contextual references, such as identifying the first hotel, directly from the native messages history.
• ▪The Claude Commerce Agents blueprint models user interface components as typed tools, such as present_products or present_itinerary, which are validated by the server before rendering.

Prompt caching optimization

• ▪Anthropic reports that cached reads in the Claude Commerce Agents blueprint cost one-tenth of fresh tokens, while cache writes carry an approximate 1.25x cost premium.
• ▪The Claude Commerce Agents blueprint structures prompt caching by ordering requests from global to session to volatile data to target cache hit rates of 90% to 99%.
• ▪The Claude Commerce Agents blueprint runs memory extraction asynchronously in a separate process, which Anthropic measured to achieve 13% higher fact recall than in-turn saving.

Multi-platform deployment options

• ▪The Claude Commerce Agents can be executed using the Messages API, the Claude Agent SDK, or Claude Managed Agents in beta.
• ▪The Claude Commerce Agents code is compatible with any Anthropic client, allowing deployment on the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI.
• ▪The Claude Commerce Agents repository runs locally on Python 3.11+ and Node 22, requiring an Anthropic API key.