Agent-led buying needs rails, and three protocols are forming them. They are easy to confuse, so here is the plain-language map of who does what and how they fit together.
ACP: the commerce flow
The Agentic Commerce Protocol, from OpenAI and Stripe, lets an AI agent select products and complete checkout programmatically, through a defined interface rather than puppeteering a store's UI. It is the piece that handles the actual buying flow: pick the item, place the order.
AP2: the payment authorization
Google's Agent Payments Protocol addresses a different question: how is an agent permitted to pay on a person's behalf, safely. Where ACP handles the commerce, AP2 handles the trust and authorization around the money. They are complementary, not competing.
MCP: the callable tools
The Model Context Protocol, from Anthropic, is the general standard for exposing callable tools to agents, including search-the-catalog and add-to-cart actions. It is the layer that lets an agent use your site as an interface instead of scraping it. ACP and AP2 are commerce-specific; MCP is the broader pattern underneath.
How they fit together
Think of it as three jobs: MCP lets the agent operate your store, ACP standardizes the purchase, and AP2 authorizes the payment. You do not have to bet on a single winner to act, because all of them depend on the same prerequisite.
What to do now
Get the foundation right before chasing any one protocol: product data, pricing, availability, and policies that a machine can read and trust, marked up with Product and Offer schema. That work pays off no matter which rail an agent uses. Then, as an early mover, track the protocols and pilot where it fits. The buyer side of all this is in AI shopping agents, and the full picture is in agentic commerce.