Anthropic has published a detailed technical guide on building effective commerce agents using Claude, based on real-world deployments across retail, travel, and telecom companies. The guide covers architecture, latency optimization, and production practices for AI agents that help customers buy and sell online more efficiently.

What Makes a Commerce Agent Different from Other AI Tools?

A commerce agent is specifically designed to simplify buying and selling across online catalogs. Unlike general-purpose chatbots, these agents handle complex, multi-step tasks like searching products, comparing options, building shopping carts, managing returns, and even handling inventory or pricing decisions for business users. Some agents face consumers directly, helping them search and assemble orders; others face the business side, answering questions about sales, running promotions, and managing inventory.

The key insight from Anthropic's guide is that the most effective commerce agents use a single unified agent architecture rather than splitting tasks across multiple specialized subagents. This matters because commerce conversations are tightly coupled, requiring shared context across multiple steps and intents. When you hand off between subagents, you lose state and context, which degrades response quality and adds latency and token costs.

How Should Teams Structure Their Commerce Agents?

• Single Agent with Skills: Rather than creating separate subagents for different domains like search, returns, or customer service, the most effective approach uses one main agent equipped with modular skill instructions that load into the system prompt as needed, avoiding the overhead of handoffs.
• System Prompt vs. Skills Decision: Instructions that the agent needs on most turns go in the system prompt; less frequent instructions become skills that load only when relevant. For commerce, product search lives in the prompt since nearly every session touches it, while features like customer care and personalization become skills.
• Tools Built on Existing Systems: Agent tools should call a company's existing search, ranking, cart, inventory, and analytics systems rather than reimplementing them, allowing the model to focus on judgment and presentation rather than core business logic.
• UI Components as Tools: Most commerce agent responses are visual components like product carousels, itineraries, or seat maps rather than plain text, so agents must emit structured schemas that render these interfaces.

In Anthropic's reference implementation, the shopping agent's system prompt holds core grounding, cart semantics, and presentation rules, while skills cover search-discovery, purchase-research, planning-goals, customer-care, and memory-personalization. The merchant-facing agent splits similarly, with skills for performance-insights, catalog-listings, inventory-operations, pricing-promotions, and marketing-campaigns.

What Results Have Enterprises Seen?

Enterprise customers using Claude-based commerce agents have observed larger shopping carts and more efficient seller operations in production deployments. Across several enterprise deployments Anthropic studied, a single agent with skills consistently outperformed both one-prompt-for-everything designs and subagent architectures on quality, often at lower cost and latency per task.

The practical implication is that teams building AI shopping experiences should prioritize architectural simplicity and unified context over the temptation to create domain-specific subagents. This approach scales better as models improve and handle longer context windows, more skills, and more tools with each generation.

How to Optimize Commerce Agents for Speed and Cost

• Minimize Latency: Reduce perceived latency by returning partial results early and streaming responses, while using prompt caching to avoid reprocessing the same instructions and context on repeated requests.
• Choose the Right Model Configuration: Select the appropriate Claude model and settings based on your latency and cost requirements; faster responses may justify using a smaller model for certain tasks.
• Reshape Tool Responses: Return only the fields the model needs to reason with, dropping unnecessary data like image URLs on every search result, and include helpful error instructions instead of generic error codes to guide the agent's next steps.

Anthropic's guide emphasizes that production commerce agents require robust evaluation practices to ship a non-deterministic system reliably. Teams need memory systems that survive user sessions, safety enforcement built into the agent harness rather than the model itself, and comprehensive evals that catch quality regressions before they reach customers.

The broader takeaway is that as large language models become more capable, the bottleneck for enterprise AI applications shifts from model capability to thoughtful system design. For commerce specifically, that means treating the agent as part of a larger ecosystem of existing business systems, not as a replacement for them.