Agentic shopping needs direct merchant APIs because browser-driving agents are slow, brittle, and often trigger fraud systems at payment.
2
MCP, A2A, ACP, UCP, and AP2 cover different parts of the flow, from tool access and agent communication to checkout and payment authorization.
3
Commerce agents need behavior, protocol, and latency evals because unexpected uses and protocol failures otherwise appear in production.
Summary
Ahnaf Prio describes what Best Buy has learned while working on agentic commerce. Shopping already makes up about 45% of sessions on major AI assistants, but the first browser-based agents struggled with screenshots, DOM changes, slow form filling, and merchant fraud checks. The emerging approach uses merchant APIs and product feeds. MCP gives an agent access to tools, A2A lets agents communicate, ACP and UCP define commerce checkout flows, and AP2 adds payment authorization details such as a spending limit and revocation URL. Prio demonstrates the complete flow with Ginny, an orange tabby turned bakery agent. The demo covers product search, cart state, checkout, payment authorization, catalog feeds, and both UCP and ACP schemas. He is direct about the unresolved parts, including payment autonomy, ACP and UCP convergence, identity, and multi-agent delegation. His practical advice is to write evals for behavior, protocol compliance, and latency before deploying.
Agentic commerce covers the whole shopping journey, not just product discovery
Prio defines agentic commerce as AI helping with a customer's shopping journey. That journey includes discovery, deciding whether an item is needed, loyalty, pricing, fulfillment, and post-fulfillment tasks. About 45% of agent sessions on major providers such as ChatGPT and Google Gemini are related to shopping. The current model still keeps a human in the loop. A more autonomous version would have an agent talk to merchants, negotiate, and make the payment. Prio frames his architecture around the current human-approved flow while showing how it could extend toward autonomous shopping.
Browser-driving shopping agents failed because they were slow and looked suspicious to merchants
Earlier shopping agents took screenshots, read the DOM, navigated merchant websites, filled forms, and handled loyalty steps. Prio says this approach was "really clunky and slow and brittle." From a merchant's perspective, an AI impersonating a browser triggers fraud alarms. That often leaves the agent stuck during payment. The newer approach avoids the browser and lets an AI assistant surface a product, call a merchant checkout API, use a payment token or mandate, and confirm the order without manually operating the merchant site.
Commerce protocols encode distinctions that shoppers usually do not see
OpenAI's ACP and Google's Universal Commerce Protocol, or UCP, provide competing commerce primitives. Prio uses a simple example: a shopper sees little difference between adding one item and adding a second quantity, while a merchant may treat the second quantity as another line item on the same SKU. Commerce APIs need to represent those details consistently. Without standardized primitives, the shopping flow stays clunky because the customer agent and merchant systems do not share the same understanding of cart and checkout operations.
Each acronym covers a different connection in the shopping system
MCP, the Model Context Protocol, gives an agent access to merchant tools such as product search, product details, and loyalty. A2A defines how agents communicate, including a customer agent talking with a merchant agent. ACP and UCP define commerce primitives and checkout behavior. AP2 is an agentic payment protocol built as an extension of UCP. Prio's mental model separates tool access, agent-to-agent messages, commerce operations, and payment authorization instead of treating all of these protocols as interchangeable.
Product feeds scale better than asking every merchant to answer every catalog search
ACP and UCP currently rely on merchants sending product feeds rather than exposing a search catalog call. Prio says the feed contains organized product information and can include updates when products change. With M merchants and N products, searching across every merchant and product creates a large call pattern. A feed can be indexed ahead of time. He also shows that Meta's feed is similar to, but different from, the other specifications, so merchants may still need to support several feed formats.
Payment flows still keep a human or payment processor responsible
Prio says current commerce systems are not yet ready for fully autonomous payment methods such as x402. ChatGPT payments use a shared payment token, while Gemini's UCP flow accepts Google Pay. The current design keeps a merchant and payment processor involved, with the processor taking responsibility for initiating the payment. AP2 adds information needed for greater autonomy: who authorized the agent, what it can buy, the maximum amount it can spend or negotiate within, a revocation URL, and proof connected to the user.
The Ginny demo connects agent messages, tools, checkout state, feeds, and payment authorization
Ginny, Prio's orange tabby, becomes a bakery merchant agent in the demo. A customer agent asks for products through A2A, then uses an MCP product-search tool to retrieve them. After adding shortbread to the cart, the flow moves through UCP checkout states: not ready for payment, ready for payment, and completed. Prio issues an AP2 token after selecting a credit or debit card. The token includes a maximum amount, currency, revocation option, and single-use setting. The demo also compares equivalent checkout calls under ACP and shows catalog synchronization on a timeline.
Agent commerce needs evals for behavior, protocol compliance, and speed
Prio compares deploying conversational commerce without evals to playing whack-a-mole. He cites a story about people using Chipotle's agent to ask programming questions, because the system had not been told to reject that behavior. Other failures could reveal discount codes or information about who else is checking out a product. He recommends behavior evals, protocol-compliance checks for feeds and integrations with systems such as GPT or Gemini, latency benchmarks, and an LLM used as a quality judge. He says every second in a retail shopping flow can give a faster site a chance to win the customer.
"Shopping has different facets to it. For instance, there's discovery, there's figuring out the aspects of do I actually truly need it, understanding and deciding, there's loyalty, there's pricing, there's fulfillment, post fulfillment."01:14
Who should watch
You are building a merchant or customer agent and need to understand how tool calls, agent communication, product data, checkout, and payments fit together.
Your shopping agent still drives websites through a browser and runs into slow forms, fraud checks, or failed payment steps.
You are integrating with ChatGPT, Gemini, or other commerce surfaces and need practical evals for behavior, protocol compliance, and latency.
