What Is a Commerce Agent? Shopping Agents vs Merchant Agents

Definition

A commerce agent is an AI agent that can assist or execute buying and selling workflows against commerce systems and product catalogs.

Unlike a chatbot that only produces language, a commerce agent can select tools, retrieve operational data, maintain workflow state, and take bounded actions. It might compare products, choose a variant, add an item to a cart, find low-stock SKUs, propose a promotion, or stage a catalog update for approval.

“Commerce agent” is a platform-independent category. Anthropic’s Claude Commerce Agents provide one current implementation of the pattern, but shopping, merchant, retail, marketplace, and procurement agents can be built on different models and stacks.

Shopping agent vs merchant agent

A shopping agent acts for the demand side. It converts buyer intent into product discovery, evaluation, selection, and potentially a transaction. Its job is not merely to return similar text; it must resolve the exact product and variant that meet the buyer’s constraints.

A merchant agent acts for the supply side. It helps a seller or operator analyze and change catalog, inventory, pricing, promotion, and merchandising workflows. Because its actions can affect many customers or products, proposed and executed writes should be separated and governed.

The shared product truth

The two agent types have different goals but depend on the same operational product layer.

A shopping agent asking whether a blue 230 V model is available and a merchant agent considering a promotion for that model must agree on:

• the canonical product, family, variant, sellable SKU, offer, and packaging level;
• decision-critical attributes and units;
• compatibility, substitution, bundle, and supersession relationships;
• current price, inventory, availability, and lifecycle state;
• shipping, return, eligibility, and restriction policies;
• provenance, confidence, validation, and freshness;
• the actions permitted for the current actor and context.

If those facts are duplicated or inconsistent across tools, the agents can each behave reasonably and still contradict one another.

How a commerce agent works

A commerce agent typically combines four elements:

1. Model and instructions: interpret intent, plan steps, and decide which authorized tool to call.
2. Tools: narrow interfaces to catalog, search, cart, order, analytics, inventory, pricing, campaign, and policy systems.
3. State and orchestration: retain the task context, tool results, approvals, and progress across steps.
4. Controls and evaluation: authenticate the actor, constrain permissions, validate inputs and outputs, require approvals, log decisions, and test outcomes.

The tool boundary matters. Core price calculation, inventory reservation, policy enforcement, and catalog validation should remain deterministic services where possible. The agent can decide when to call them and reason over their results without being asked to recreate their business logic from prose.

Examples

Shopping-agent examples

• ask clarifying questions when a technical constraint is missing;
• compare products using normalized attributes;
• select a compatible accessory for an exact variant;
• find an approved substitute for an unavailable item;
• calculate the correct number of cases for a requested quantity;
• add a revalidated SKU and quantity to a cart.

Merchant-agent examples

• find listings with stale or conflicting product facts;
• identify low-stock variants using lead time and demand;
• propose price changes within a margin policy;
• select promotion candidates subject to eligibility rules;
• stage evidence-backed attribute corrections;
• execute approved changes and preserve a reversible audit trail.

Related terms

Agentic commerce is the broader mode of commerce in which AI agents participate in discovery, decision, transaction, and operations. A commerce agent is an actor within that environment.

Shopping assistant often describes a conversational discovery interface. It becomes an agent when it can plan and use tools to progress a workflow, especially when it can change state.

Recommendation engine ranks items from a defined input and objective. A commerce agent may call one, but also clarifies intent, uses other systems, and coordinates multiple steps.

Merchant copilot usually suggests work for a human. A merchant agent can also take bounded actions. The practical distinction is its tool access and authority, not whether the interface uses the word “copilot.”

Procurement agent acts for an organizational buyer and may include supplier selection, approval, policy, and purchase-order workflows. It overlaps with the shopping-agent side but usually operates under enterprise controls.

Data requirements

A commerce agent needs more than searchable copy. Its tools should expose:

Why catalogs fail commerce agents

Catalogs built for pages and feeds often tolerate ambiguity because a person can interpret it. Agents turn that ambiguity into operational errors. Common examples include a family page without a variant mapping, 12 without a unit, package dimensions stored as product dimensions, a compatibility claim hidden in prose, stale inventory copied into search, or two suppliers creating duplicate versions of one product.

The remedy is not to add more text to the prompt. It is to establish an operational product-data process: identify, assemble, validate, expose, control, and learn.

Current example: Claude Commerce Agents

Anthropic’s commerce-agent pattern distinguishes a buyer-facing shopping agent from a seller-facing merchant agent and connects them to existing commerce systems. That separation is useful beyond any one platform because it makes the authority boundary visible: the model coordinates work, while merchants remain responsible for the systems, product truth, policies, and approvals behind the tools.

Read what a product catalog must provide before connecting Claude, follow the catalog preparation guide, or run the 20-case catalog test.

FAQ