Anatomy of Effective Commerce Agents: A Production Guide
Anthropic distills a year of building Claude-based commerce agents with retailers, travel, and telecom customers into a production playbook. It argues against intent routers and domain-split subagents: commerce sessions are tightly coupled across intents, and handoffs lose state, add tokens, and increase latency. Instead, one agent in a standard loop with skills for the long tail outperforms both one-prompt and subagent designs. UI output should be modeled as presentation tools — typed tool calls, not client-parsed custom tags — at the cost of streaming granularity. Latency is attacked on end-to-end and perceived fronts; prefix caching with a global/session/volatile layout can sustain 90–99% hit rates. Memory facts are written asynchronously by a separate extractor into your database (13% higher recall in their commerce eval), then read in three layers. Safety is enforced in the harness: model-originated writes are only staged and need approval, only server-issued IDs are accepted, and third-party content is sanitized. Evals are snapshots, not simulated-user conversations, with paired positive and negative cases. An open-source reference implementation, anthropics/commerce-agents, is included.