The Shift from Browser Automation to Standardized Protocols

Early attempts at agentic shopping relied on "screenshot and DOM" agents that navigated websites like humans. These were brittle, slow, and frequently triggered fraud alarms. The industry is now moving toward standardized commerce primitives that allow agents to communicate directly with merchant APIs.

Key protocols currently defining this space include:

  • MCP (Model Context Protocol): Standardizes how agents identify and access tools (e.g., product search, loyalty checks).
  • A2A (Agent-to-Agent): A specification for communication between customer-facing agents and merchant-side agents.
  • ACP (OpenAI) & UCP (Google): Competing commerce primitives that define how product data and checkout flows are structured.
  • AP2 (Agentic Payment Protocol): An emerging extension for autonomous payments that includes an authorizing party, spend ceilings, and revocation URLs to manage liability.

Instead of agents performing live searches against a merchant's catalog—which does not scale as the number of merchants and products grows—the current best practice is for merchants to push product feeds to the AI providers. This allows the AI to index inventory ahead of time, ensuring faster response times and better integration with retail media and sponsored product ranking systems.

The Critical Role of Evals in Production

Building agentic commerce without rigorous evaluation is described as "playing whack-a-mole." Because agents can be manipulated (e.g., users asking a food-ordering agent for programming help), developers must implement specific testing frameworks:

  • Behavioral Evals: Ensuring the agent stays within its domain and does not leak sensitive information or offer unauthorized discounts.
  • Protocol Compliance: Validating that product feeds and API calls conform to the specific schemas required by platforms like ChatGPT or Gemini.
  • Latency Benchmarks: Monitoring response times, as retail conversion drops significantly if the shopping flow feels sluggish compared to traditional web interfaces.
  • LLM-as-a-Judge: Using an LLM to evaluate the quality and correctness of agent responses against predefined success criteria.