Anthropic Claude Commerce Agents Open-Source

Anthropic released an open-source reference implementation for e-commerce AI agents, including Shopping and Merchant agents, with detailed architecture and safety guidance. - Two agents: Shopping Agent (consumer-facing, embedded in apps/sites) and Merchant Agent (for store operators). Four vertical demos: retail, travel, telecom, ticketing. Three run modes: Messages API, Claude Agent SDK, Managed Agents (beta). - A Claude Code plugin 'commerce-builder' provides commands to scaffold, add flows, author evaluations, and review agents. - Architecture: single agent + skills, not multi-subagent, because e-commerce dialogues are tightly coupled; subagents cause lossy state handoffs and higher latency/cost. Skills cover search, purchase research, planning, customer care, memory. - System prompt holds security, legal, brand constraints, and high-frequency instructions; skills handle long-tail intents. Tools call existing backend systems (search, inventory, promotions) rather than recreating logic. - UI components are defined as tools (e.g., present_products) to avoid custom markup and enable native replay. - Latency optimization: fewer rounds via preloaded context and parallel tool calls; prompt caching targets 90-99% hit rate by ordering requests from static to volatile. - Memory stored in enterprise database, asynchronously extracted after each turn (13% higher recall than model-called saves), with three-tier read and user control. - Safety: enforcement in harness, not prompt; model only proposes, human or policy executes; write operations use server-signed IDs; purchase limits serialized; third-party content sanitized; regulated fee disclosures verbatim. - Evaluation: snapshot-based (construct state + user message + run + assess final state and rendering); 50-100 cases per flow, each positive paired with a negative; injection attacks tested from user messages and data plane. - Reported results: retailers using Shopping Agent saw cart value increase up to 35% and purchase completion likelihood up 60%.
Key insights
Single-Agent Architecture with Skills Outperforms Multi-Agent Setups for Commerce
Anthropic's design explicitly rejects subagent orchestration for e-commerce, arguing that shopping dialogues require shared state (cart, preferences, history) and that handoffs between subagents introduce lossy state transfer, higher token usage, and latency. Their single-agent + skills approach consistently outperformed both monolithic prompts and multi-agent systems in enterprise deployments.
Safety Is Enforced at the Harness Level, Not in Prompts
The reference implementation places all safety enforcement outside the model: the agent can only propose actions; write operations generate staged changes with server-signed IDs that require approval; purchase limits are serialized and validated server-side; third-party content is sanitized; and regulated fee disclosures are served verbatim from the backend. This design prevents prompt injection and model errors from causing harmful transactions.
Prompt Caching and Asynchronous Memory Are Key Cost and Performance Levers
The system targets 90-99% prompt cache hit rates by ordering request segments from static (system prompt, tool definitions) to volatile (time, current page). Memory extraction runs asynchronously after each turn, improving fact recall by 13% over model-called saves without adding conversation latency. These optimizations make the agents economically viable at scale.
Open-Source Release Includes Production-Grade Evaluation and Organization Patterns
The repository provides not just agent code but a full evaluation methodology: snapshot-based testing, positive/negative case pairs, injection attack tests, and 50-100 cases per flow co-authored with domain experts. It also outlines organizational practices like ownership-follows-system, canary releases, and CI that runs core plus affected test sets, making it a blueprint for enterprise adoption.
Reported Business Impact Shows Significant Conversion Uplift
Anthropic claims that retailers using the Shopping Agent saw cart values increase by up to 35% and purchase completion rates rise by 60%. These figures, if validated, suggest that well-designed commerce agents can directly impact revenue, moving beyond cost savings to top-line growth.