Commerce agent architecture

Paths are in the reference repo: commerce_common/ is commerce-common/commerce_common/, shopping_agent/ is shopping-agent/core/shopping_agent/, merchant_agent/ is merchant-agent/core/merchant_agent/, and the runtimes are shopping-agent/runtime-messages-api/shopping_agent_runtime/ and merchant-agent/runtime-messages-api/merchant_agent_runtime/.

One loop

• One model owns the conversation. A turn is ShoppingAgent.stream_turn or MerchantAgent.stream_turn (each runtime's orchestrator.py): the model reads the cached prompt, calls tools (one round's calls run concurrently), and ends with text plus presentation calls. There is no router, classifier, or hand-off.
• Every call on every path runs through the role executor (ShoppingToolExecutor in shopping_agent/executor.py, MerchantToolExecutor in merchant_agent/executor.py) over BaseToolExecutor in commerce_common/execution.py. The Messages API runtime, the Agent SDK toolset, and the MCP server call the same execute, which never raises; after max_tool_iterations rounds (commerce_common/config.py) the runtime forces a round without tools.

Where a rule lives

A rule the core journey needs on most conversations moves a layer down. This table is the one statement of the layering.

Skills

• A skill is a directory holding SKILL.md: frontmatter name and description, then the body (parse_skill_md in commerce_common/skills.py). The description names the request class; it carries no sample utterances.
• The static prompt carries the index alone (SkillRegistry.index_block, sorted by name) and load_skill returns one body (_load_skill in commerce_common/execution.py). On the Agent SDK, ensure_project_skills links the same directories into .claude/skills/ and SKILL_TOOL_ADAPTER points the model at the SDK's Skill tool (commerce_common/agent_sdk.py); the hosted manifests list the directories under skills[] in agent.yaml.

The backend interface

A deployment implements one abstract class per role; nothing else reaches its systems.

• StorefrontBackend (shopping_agent/backend.py): 11 abstract methods, catalog (2), cart (4), preferences (1), orders (2), policies (1), fulfillment (1); get_account_context, get_disclosure, and checkout_handoff have defaults. No method places an order or moves money; checkout_handoff names where payment happens (a hosted checkout URL, or one per seller) and the host renders it.
• MerchantBackend (merchant_agent/backend.py): 16 abstract methods, reads (8: performance 3, catalog 2, inventory and order health 2, pricing 1) and the change lifecycle (8: five stage_*, get_pending_changes, apply_change, discard_change); execute_analysis_query, get_analysis_schema, get_merchant_context have defaults.
• A product or listing sold by size, color, or tier is a family record carrying options, its variants each a record with option_values and variant_of; the cart and price or restock writes take a variant's id and the gates hold a family's. docs/backends.md has the mapping from common catalog models and what to return for a figure the platform cannot supply.
• Eligibility, pricing, inventory, and quantity rules are enforced in the backend; the gates check provenance and caps (commerce-trust-safety). Everything a backend returns is fenced before the model reads it.
• A system that is not wired yet is a method that fails (merchant methods raise ChangeNotApplicable from merchant_agent/changes.py); the tool stays registered, so prompt bytes do not change. A system the business does not have at all is absent: an enable_* switch on the role's config turned off, which removes its tools (absent_tools()), prompt lines, and grounding rule on all three paths, and the executor refuses those names; the skills that need it are parked under skills/_staged/, and the bytes are then fixed for that deployment. NotOffered (shopping_agent/backend.py) is the narrower signal: one item or seller the store does not serve while the system itself is on.
• executor_class= on ShoppingAgent / MerchantAgent, the SDK toolsets, and the MCP servers' build_server takes a subclass of the role's executor for a deployment's own domain_error mapping or result wording.

Delegates

A second model call takes one shape: a DelegateExtension (commerce_common/delegation.py). It receives a brief and the handles in DelegationContext, never the conversation or the executor; it returns one result validated against result_model; it cannot write, present, or call a delegate; _run_delegate in commerce_common/execution.py caps its calls per turn at max_delegate_calls_per_turn. MerchantAgent takes extra_delegates, and its built-in instance is run_analysis (commerce-merchant-operations). ShoppingAgent registers none.

Model fields

model runs the turn loop and memory_model the post-turn extraction (BaseAgentConfig in commerce_common/config.py); analysis_model runs the delegate, and None means model (MerchantAgentConfig in merchant_agent/config.py). The SDK runtimes copy config.model into the options; the manifests set model:. Choose the values with your own evals.

Do not

• Vary tools[] or the static prompt by request (commerce-prompt-caching).
• Put a most-turns rule in a skill, or a one-tool rule in the prompt.
• Enforce a cap or a permission in prompt text alone; it belongs in the executor or the backend.
• Let the model author a price, a figure, or a term; components are joined from server records (commerce-ui-tools).
• Add a domain tool to a core package; a vertical adds UI through PresentationExtension and keeps the rest in its own code.