Anthropic Releases Open-Source Claude Commerce Agents Blueprint

The shopping agent is built to run inside a merchant's app, handling catalog searches, multi-item requests, option comparisons, cart building, and order and returns questions in one conversation. Its five skills are search-discovery, purchase-research, planning-goals, customer-care and memory-personalization. The merchant agent answers sales-performance questions, raises inventory alerts, gives pricing and promotion recommendations, and drafts campaigns, using skills such as performance-insights, catalog-listings, inventory-operations, pricing-promotions and marketing-campaigns. Both agents can run through three execution paths—the Messages API, the Claude Agent SDK, or Claude Managed Agents in beta—from a single definition of prompts, skills, tool contracts and gates. A Claude Code plugin, commerce-builder, can scaffold a new agent or review an existing one.

Anthropic's architectural claim, detailed in the deep-dive, argues against building an intent router and against assigning one subagent per domain. Commerce sessions are tightly coupled conversations, and handoffs between subagents are state-lossy: an orchestrator holds the cart, preferences and history, and each transfer can raise token use and latency. Agent skills deliver modularity without that tax because skill instructions are loaded into the agent that already holds the conversation history. Across several enterprise deployments, Anthropic says a single agent with skills beat both one-big-prompt designs and subagent designs on quality, often at lower cost and latency. Subagents still make sense for narrow, self-contained tasks such as deep research. The blueprint places roughly a third or more of frequent traffic in the system prompt and the rest in skills; safety rules, brand constraints and key user facts always remain in the prompt.

The blueprint treats most commerce responses as UI components rather than raw prose. Instead of prompting the model to emit custom tags, it exposes each component as a typed tool—such as present_products, present_itinerary and present_plan_comparison—and the server validates arguments before the client renders them. Because these tool calls appear natively in the messages array, reloading history needs no custom parser, and the agent can refer to objects from earlier presentation calls. Token-level streaming is enabled with eager input streaming to avoid server-side buffering.

Anthropic addresses latency by distinguishing end-to-end latency from perceived latency, streaming components as they form and showing plain-language progress lines. Since a rendered response may run 500–700 output tokens, that makes a difference. Eager tool dispatch—executing each call as its arguments finish streaming—cuts multi-second pauses to a few hundred milliseconds. Prompt caching is the main cost lever: requests are ordered global, session, then volatile, because caching is prefix-based and a volatile value at the top would break the cache. Well-tuned deployments can reach 90–99% cached-read hit rates. Memory extraction runs asynchronously in a separate process, which Anthropic measured at 13% higher fact recall than an in-turn save tool. Money, writes and IDs remain gated in code, meaning the model proposes operations and the harness applies them.