E-commerce is shifting from conversational AI to Agent-to-Agent: the buyer's agent talks directly to the merchant's, no website in the middle. What makes this possible isn't a single new protocol — it's an evolution of MCP. MCP defined the universal contract for agent-to-tool communication. A2A built on its primitives for agent-to-agent. UCP took both and specialized them for the commerce contract — discovery, cart, checkout, identity, orders. ACP took a parallel path on the OpenAI side. AP2 extended into agent-initiated payments. Each layer inherits MCP's design principles and adapts them to a narrower problem.
This talk traces that evolution from a platform vendor's seat. At Brainform.ai we build the agentic layer between merchants and AI agents — the place where MCP's commerce descendants meet in practice. I'll walk through what each protocol inherited from MCP, where they diverge, and what platform builders need to layer on top. We'll trace a single shopping interaction across all four, look at what's already in production and what's still vendor demos, and discuss the design decisions that fall on platform builders today — identity model, fraud and returns, the agent-of-record problem, and what happens when one agent in the chain misbehaves.
We'll cover the failures that already happened in 2026 (OpenAI killing Instant Checkout after the conversion drop) and the open questions still on the table: governance, dispute resolution, who pays the agent's fees, how merchants stay discoverable. These aren't protocol bugs — they're the boundary conditions emerging when commerce moves from human-driven to agent-driven.
Best fit for MCP server and client builders working in commerce, architects designing multi-protocol agent systems, and anyone curious about how MCP grows into specialized verticals.