Anthropic publishes a blueprint for its commerce agents

Anthropic has released an Apache 2.0 blueprint for building Claude agents that support shopping and merchant operations without directly handling payments or sensitive changes.

An assistant can search for a product, compare several offers, suggest an alternative, and prepare a shopping cart. But it must stop before payment. In merchant-facing tools, it can also recommend a price reduction or promotional campaign, though a person must still approve the change.

These boundaries define Anthropic’s new guide to building commerce agents. Titled A guide to the anatomy of effective commerce agents, it accompanies an Apache 2.0-licensed blueprint and presents a reference architecture for developing services powered by Claude.

This is neither a new model nor a turnkey commerce platform. Anthropic is providing code, sample interfaces, tools, and evaluations that each company must adapt to its own catalog, policies, and internal systems.

The Commerce Agents repository contains two categories of agents. The first is designed for consumers. It can search for offers, compare options, answer questions about orders or store policies, remember certain preferences, and prepare the elements required for a purchase.

The second is intended for merchant employees. It analyzes sales, checks inventory, prepares product listings, recommends prices or promotions, and drafts campaigns. The goal is not so much to replace existing management software as to provide a conversational interface capable of coordinating several operations.

Anthropic applies this architecture to four industries: retail, travel, telecommunications, and ticketing. Each vertical includes a customer-facing application and another for internal operations, for a total of eight demos.

The data, brands, and companies used in these examples are fictional. The code does not place real orders, charge payment cards, or modify live listings. The demos also lack a production-ready authentication system.

That distinction matters. A company adopting the blueprint must still connect its catalog, search engine, inventory, order management, shopping cart, payment system, pricing, and promotional tools. The repository does not ship with business connectors.

Anthropic recommends placing the agent on top of the systems a company already uses. A pricing tool, for example, should query the system that actually owns the pricing rules instead of attempting to reproduce them inside an instruction sent to the model. The same principle applies to availability, discounts, refunds, and order limits.

The blueprint can run with the Messages API, Claude Agent SDK, or Managed Agents. It also includes a Claude Code plugin with commands for creating a new implementation, adding a user flow, writing evaluations, or reviewing an existing agent.

Anthropic’s recommended architecture centers on one primary agent supported by tools, modular skills, and an evaluation suite. The company advises against automatically creating a separate subagent for search, shopping carts, orders, promotions, and every other commerce function.

According to Anthropic, handoffs between subagents can lose context, consume more tokens, and add several seconds to some customer journeys. Commerce domains also overlap: a substitution request may require simultaneous access to the catalog, inventory, customer preferences, and promotional rules.

Anthropic says a single agent equipped with specialized skills performed better than either a very long instruction set or a multi-subagent architecture across several enterprise deployments. However, it provides no public protocol, dataset, or detailed results that would allow this advantage to be independently measured.

Subagents are not excluded entirely. The guide reserves them for lengthy research tasks that can be isolated from the main conversation or for transfers to genuinely specialized services, particularly when regulatory requirements justify a clear separation.

Within the primary agent, frequently used rules should appear in the system prompt. Less common procedures can be loaded as skills when the conversation requires them. Anthropic offers a rough threshold: information relevant to about one-third of requests generally belongs in the permanent context.

Safety instructions, legal requirements, brand rules, and essential user information should nevertheless remain available at all times, even when they apply to only a small share of interactions. Their importance takes precedence over how often they are used.

Tools are not limited to retrieving or modifying data. They can also describe structured interface components such as product carousels, itineraries, seating charts, or graphs. The server receives an output that follows a defined schema, validates and enriches it when necessary, and then instructs the browser to display it.

This approach avoids giving the model unrestricted control over interface generation. It also makes it possible to stream a component before the full written response is complete. Early delivery can improve perceived speed, but it complicates validation because the first elements arrive before the complete structure is known. Anthropic therefore recommends implementing retries when the final output fails to match the expected schema.

Response time is treated as a property of the entire user journey. A single interaction may involve several model calls, multiple service requests, and several rendering stages. Optimizing the initial response alone offers little value if the agent then creates unnecessary steps.

The guide recommends loading information that is likely to be needed from the start, running independent calls in parallel, and improving the slowest internal services. It also suggests that a more capable model may sometimes reduce total time and cost if it completes the task in fewer steps.

Intermediate results also influence how fast the system feels. A product or itinerary can appear while the agent continues working, accompanied by a brief message explaining what it is doing. The user is no longer left waiting in front of an inactive interface.

Prompt caching is presented as the largest source of cost savings.