The Shift from Browsing to Structured Commerce
Traditional web browsing is inefficient for AI agents, which struggle to parse HTML and consume excessive tokens. To enable agentic commerce, merchants must adopt a structured approach. Instead of relying on standard web interfaces, merchants should publish a merchant-capabilities-manifest (a JSON file in the /.well-known/ directory) that declares supported payment methods and API endpoints. Catalogs and policies (shipping, returns) must also be exposed as structured JSON. This allows agents to filter, rank, and justify recommendations based on specific attributes rather than hallucinating or failing to find information.
Designing Agent Personas and Guardrails
An agent's behavior is defined by its system prompt (persona) and its instruction loop. A poorly designed persona—such as an "aggressive salesman"—can lead to manipulative behavior and dark patterns. Developers must implement a strict guardrail checklist to maintain user trust:
- Transparency: Always disclose that the user is interacting with an AI.
- Control: Honor "stop" and "cancel" commands immediately.
- Financial Safety: Enforce hard caps on transaction totals and disclose all fees upfront.
- Accountability: Log all agent decisions and attribute matches to provide an audit trail of why a specific product was recommended.
Securing Transactions with Shared Payment Tokens
Security in agentic commerce relies on abstracting sensitive data. Rather than passing raw credit card numbers, agents should utilize "shared payment tokens." In this model, the agent receives a token from a payment provider (like Stripe) and passes it to the merchant. The merchant unwraps only the necessary data to process the transaction, and the payment provider enforces limits and validates the request. This ensures that neither the agent nor the merchant ever handles raw payment credentials, significantly reducing the security surface area for the user.