Anthropic Released Claude Commerce Agents: An Apache-2.0 Blueprint for Shopping and Merchant Agents Across Retail, Travel, Telecom and Entertainment Anthropic发布Claude商业智能体：面向零售、旅游、电信和娱乐领域的Apache-2.0购物与商家智能体蓝图

TL;DR

• Anthropic released anthropics/commerce-agents as an Apache 2.0 blueprint containing a shopping agent and a merchant agent, with four runnable verticals: retail, travel, telecom, and entertainment.
• The architecture advocates a single agent loop with modular skills over subagent handoffs or monolithic prompts, claiming better quality, lower cost, and reduced latency.
• UI components are implemented as typed tools (e.g., present_products, present_itinerary) so that history remains natively parsable and layout references are resolvable.
• Prompt caching is a primary cost lever, with best deployments achieving 90–99% hit rates by ordering prompts global → session → volatile and avoiding volatile data like timestamps at the top.
• The blueprint is deployment-agnostic, running on Claude API, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI, with a Claude Code plugin (commerce-builder) for scaffolding and reviewing agents.

Why It Matters

Anthropic is effectively open-sourcing the architectural patterns that most teams reinvent when building commerce agents, reducing time-to-deployment and providing empirically validated design choices. The skills-over-subagents argument and the typed-tool UI pattern offer transferable lessons for any agentic system that needs to balance modularity with state continuity and low latency.

Technical Details

• Two agent types: The shopping agent (five skills: search-discovery, purchase-research, planning-goals, customer-care, memory-personalization) operates inside a merchant app over a StorefrontBackend. The merchant agent (five skills: performance-insights, catalog-listings, inventory-operations, pricing-promotions, marketing-campaigns) supports store staff over a MerchantBackend.
• Three runtime modes: Both agents run via the Messages API, the Claude Agent SDK, and Claude Managed Agents (beta), all from a single definition of prompts, skills, tool contracts, and gates.
• Skills vs. subagents: Anthropic argues that handoffs between subagents are state-lossy and expensive in tokens and latency. Agent skills load instructions into the same agent that already holds conversation history, achieving modularity without the tax. Subagents are reserved for narrow, self-contained tasks like deep research.
• UI components as tools: Commerce responses are rendered as typed tool calls rather than custom-tagged prose. This keeps component data in the messages array natively, enabling history reload without custom parsers and allowing the agent to resolve references like "the first hotel" from prior presentation calls.
• Latency and caching optimizations: Eager tool dispatch executes calls as arguments finish streaming, cutting multi-second gaps to hundreds of milliseconds. Prompt caching is prioritized as the main cost lever, with volatile data placed last in the prompt prefix. Memory extraction runs asynchronously, yielding 13% higher fact recall than in-turn saves.
• Safety gates: Financial writes, ID mutations, and other sensitive operations are gated in code—the model proposes, the harness enforces.

Industry Insight

• The skills-over-subagents pattern should be a default consideration for any agentic system where conversation state continuity is critical; the token and latency overhead of handoffs is a real but often underestimated cost.
• Treating UI components as typed tools rather than prompting for structured text is a practical pattern that generalizes beyond commerce—it ensures parseable history and enables richer client-side rendering without custom serialization logic.
• Prompt caching strategy (ordering by stability: global → session → volatile) is a high-ROI optimization that most teams overlook; achieving 90–99% hit rates can dramatically reduce inference costs at scale.

TL;DR

• Anthropic发布Apache 2.0开源的commerce-agents参考蓝图，包含购物代理和商家代理，覆盖零售、旅游、电信、娱乐四个垂直领域
• 提出"Agent Skills"架构替代传统的意图路由和子代理设计，单代理+技能模式在质量、成本和延迟上均优于子代理方案
• UI组件作为类型化工具实现，使对话历史保持原生结构，支持"第一个酒店"等指代解析
• 通过提示词缓存（90-99%命中率）、急切工具分发和异步记忆提取实现延迟与成本优化

为什么值得看

本文为电商AI代理提供了可落地的工程蓝图，解决了多代理架构中的状态丢失和延迟问题，对构建购物助手或商家运营工具的团队具有直接参考价值。其"技能而非子代理"的架构理念可迁移至其他复杂业务场景。

技术解析

双代理架构：购物代理内置于商家应用，具备搜索发现、购买研究、规划目标、客户关怀、记忆个性化五项技能；商家代理面向店员，提供绩效洞察、目录管理、库存操作、定价促销、营销战役五项技能。两者均通过StorefrontBackend/MerchantBackend对接业务系统。

Agent Skills设计：Anthropic反对意图路由和每领域子代理方案，认为商业会话是紧密耦合的对话，每次交接都会丢失状态并增加token成本和延迟。技能指令加载到已持有历史的代理中，实现模块化而无需状态转移开销。

UI组件即工具：present_products、present_itinerary、present_plan_comparison等组件作为类型化工具实现，参数由服务端验证后客户端渲染。工具调用原生存在于messages数组，无需自定义解析器即可重建历史。

性能优化策略：端到端延迟与感知延迟分离，组件流式渲染并显示进度提示；急切工具分发将多秒延迟降至数百毫秒；提示词缓存按全局→会话→易变顺序排列，缓存读取成本仅为新token的1/10；记忆提取异步运行，事实召回率提升13%。

行业启示

架构范式转移：多代理协作的"子代理路由"模式可能并非最优解，单代理+技能模块化为复杂业务场景提供了更高效的替代方案，值得在金融、医疗等领域验证。

工程化落地加速：开源蓝图降低了电商AI代理的开发门槛，团队可直接基于此构建而非从零搭建，同时支持Claude API、Bedrock、Foundry、Vertex AI多平台部署。

成本与体验平衡：提示词缓存策略和流式渲染方案展示了如何在保证响应质量的同时控制成本，为AI产品商业化提供了可复用的性能优化模板。