Stop Building Complex Subagents for Commerce

Anthropic has released a comprehensive engineering guide on building effective commerce agents using Claude, focusing on the harsh realities of enterprise deployment, latency management, and cost control. Rather than relying on complex hierarchies of domain-specific subagents, the architecture proposes a single model running in a standard agent loop equipped with modular skills and native tool calls.

• Ditch the subagent handoffs: Handoffs between specialized subagents destroy conversation context, spike latency, and inflate token costs, whereas single agents with modular skills handle multi-intent sessions far more effectively.
• Treat UI components as tools: Forcing models to emit custom HTML tags creates fragile parsers and context bloat; defining UI layouts as typed tool schemas guarantees reliable, well-formed data rendering.
• Enforce safety in the harness: Never let the model move money or execute writes directly; all high-stakes actions must be staged through server-controlled maker-checker validation flows.

Read the original: A guide to the anatomy of effective commerce agents