The Agentic Commerce Stack: Primitives and Protocols

The Shift from Browser Automation to API-Driven Commerce

Early attempts at agentic shopping relied on browser-automation agents that captured screenshots and interacted with the DOM. These methods were brittle, slow, and frequently triggered fraud detection systems. Modern agentic commerce replaces these with standardized protocols that allow customer agents to communicate directly with merchant APIs. The current architecture relies on a stack of primitives: MCP (Model Context Protocol) for tool access, A2A (Agent-to-Agent) for inter-agent communication, and commerce-specific primitives like ACP (OpenAI) and UCP (Google) for product data and checkout state management.

Standardizing Product Feeds and Payments

Merchants should expose product data via standardized feeds rather than relying on LLMs to crawl catalog search pages. This approach improves performance and allows for better inventory synchronization. For payments, the industry is moving toward scoped mandates. While current implementations often use delegated tokens or platform-specific wallets (like Google Pay), the AP2 (Agentic Payment Protocol) specification provides a more robust framework. AP2 enables the issuance of payment tokens that include specific authorizing parties, spend ceilings, and revocation URLs, allowing for safer, more autonomous checkout experiences.

The Necessity of Evals in Production

Deploying agentic commerce without rigorous evaluation leads to unpredictable behavior, such as agents answering programming questions or leaking sensitive information. Developers must implement three core types of evals: behavior evals, protocol compliance (to ensure feeds conform to ACP/UCP standards), and latency benchmarks. Using an LLM as a quality judge is a recommended approach for maintaining consistency. Without these safeguards, production environments become a game of whack-a-mole.