45% of Agent Sessions Are Shopping: Best Buy's Agentic Commerce Stack

About 45% of all agent sessions inside major providers like ChatGPT and Google Gemini are shopping-related, according to Ahnaf Prio, a senior engineering manager at Best Buy who has spent the last year building agentic commerce. By his numbers, shopping is already most of what people do with chat assistants. His AI Engineer talk maps the protocol stack that makes those purchases actually complete — and argues the screenshot-and-fill-the-form approach everyone tried first was never going to get there.

The Agentic Commerce StackAhnaf Prio, Best Buy · AI Engineer · Watch on YouTube

Why the browser-driving era didn’t work

Prio was blunt about the first generation of shopping agents — the ones that take screenshots, read the DOM, and fill forms on a merchant site for you. “Kind of just didn’t work as expected,” he said. “It was really clunky and slow and brittle.”

The deeper problem wasn’t latency, it was the other side of the connection. “If you are a merchant who’s trying to sell stuff, any engineering department of that merchant will tell you an AI impersonating or your browser is just firing up all the alarm bells,” he said. The failure point was usually payment: the flow stalls because the merchant’s fraud stack does not want an agent submitting that order.

The API-first version works today. Prio pointed to ChatGPT shopping and Google AI Mode — the surface we looked at when Google extended AI Mode into hotel booking and flight tracking — plus Instagram and Facebook moving in, a GoPuff and Grok app, and Microsoft Copilot announcing the day before his talk that UK users can buy Ray-Bans inside Copilot. The clean flow: you ask for a product, the AI surfaces it, the agent calls the merchant checkout API, no browser involved, payment flows through a scoped mandate or delegated token, and “human kind of didn’t have to touch the cart.”

The acronym stack, decoded

“One day we’re talking about MCPs, another day A2A, ACP, UCP, AP2. It’s like, what is even real?” Prio joked that if someone told him tomorrow they had invented HYPE, he would probably believe it. His mental model separates the layers by job:

MCP is how the agent discovers and calls capabilities — what products exist, a product’s details, what the loyalty program offers. “The only way to get access to the specific capabilities is through MCP tool calls.” Weighing that against plainer alternatives runs into the same tradeoffs as MCP vs. CLI for AI agents.

A2A is how agents talk to each other — domain-level agents for payments or loyalty need a standard way to message, and the same spec covers a customer agent talking to a merchant agent.

ACP and UCP are the commerce primitives, OpenAI’s attempt and Google’s. Why standardize something as mundane as a cart? Because the shopper’s view and the merchant’s diverge immediately. “For some of you who are shopping on the other side as the customer, there is not much of a difference between adding an item to cart, adding a second quantity,” he said. “But to us merchants, that’s a second line item, buddy. That’s not the same scale.”

AP2 is the agentic payment protocol — Google’s open specification, an extension of UCP, built around a scoped mandate answering: who authorized the agent, what can it buy, what’s the maximum amount, where’s the revocation URL, where’s the proof of consent.

Merchants push feeds, they don’t expose search

You would assume the product layer is a search API. It isn’t. Both ACP and UCP want merchants to push a product feed, and neither supports a search-catalog call.

Prio gave business reasons first — sponsored products, retail media, ranking — but the load-bearing argument was throughput. With M merchants and N products, live search per query means an enormous fan-out. “While if you send the product feed ahead of time, we can index that and be ready to offload when you ask for something.” His demo re-syncs the catalog every couple of seconds.

The catch is that the formats disagree. Showing Meta’s feed alongside ACP and UCP: “As you can see, they’re similar, but still different. Everyone has an opinion.” Three specs for one catalog — which is why his starter template converts to all of them.

Payments are the most conservative layer

Nothing in production is autonomous yet, and Prio framed that as a confidence problem rather than a technical one. “None of them are supporting that more autonomous form of, you know, X402 or some other kind of payments. We’re just not there yet. We’re just not confident yet.” The industry wants a human in the loop and a payment processor willing to take the liability of initiating payment.

Concretely: ChatGPT payments happen only through a shared payment token, and Gemini’s UCP path accepts only Google Pay. AP2 — the road Prio built his demo on — encodes the guardrails as data instead. His live token carried a maximum amount, a currency, a revocation path, and single-use semantics. That max-amount field is exactly what an autonomous negotiating agent would need.

The cat bakery demo

Prio built a demo starring his orange tabby Ginny, recast as a bakery agent selling baked goods, running on a Cerebras-hosted model at 3,000 tokens per second so the flow would move fast on stage.

Asking for products triggered an A2A call from the customer agent to the merchant agent, which resolved into an MCP product_search tool call — the agent inferring the tool from intent rather than being told. Adding shortbread to the cart moved through UCP’s checkout states, which Prio enumerated as not ready for payment, ready for payment, and completed. Selecting a card issued the AP2 mandate and the session flipped to complete. He then re-ran it against ACP: structurally similar calls, differently shaped.

The most instructive moment failed on purpose: Prio asked Ginny to hand over a discount code, and she refused — which is the whole reason the next section exists.

Without evals, you’re playing whack-a-mole

“Working with AI and conversational experiences without evals is playing whack-a-mole,” Prio said. A merchant agent has pricing authority, so its failure modes are commercial, not just embarrassing — a leaked discount code, or an agent revealing who else is checking out a product.

He told a Chipotle story but explicitly hedged it: “I don’t know if it’s true or not, but I found it really funny.” The claim is that when Chipotle rolled out an agent, people used it to ask programming questions — “hands-down one of the most creative way to get free AI usage when you don’t want to pay for that cloud subscription.” True or not, the lesson holds: unconstrained agents get used in unconstrained ways.

His recommended categories were specific. Behavior evals for the discount-code class of problem. Protocol compliance, because if your feed doesn’t conform, ChatGPT and Gemini won’t support it. Latency benchmarks, because “every second in retail on the shopping journey where you’re actually not selling, there are chances that the other website’s going to be faster, and people are just going to move away.” And LLM as a quality judge — nothing fancy, he said; talk to your product colleagues and write out the best use cases. Our evals guide for teams covers that groundwork.

What happens next

Prio closed with an honest split. Stable today: MCP is widely adopted, A2A is widely used, ACP and UCP are out there. Still forming: AP2 and its actual usage, whether ACP and UCP converge or merchants keep shipping two specs forever, identity and consent standards, and multi-agent checkout delegation.

His bet is on the trajectory rather than any single spec. Today is human-in-the-loop; the state he sketched is autonomous shopping, where your agent talks to merchants, negotiates — he grew up in Bangladesh, he noted, where haggling is standard — and pays.

For builders skipping external marketplaces, his advice was to adopt the primitives anyway: a merchant agent built from standardized pieces can sell through ChatGPT and Gemini later. He released the presentation plus a three-service starter template, eval templates, the ACP/UCP/Meta catalog sync, and agent skills for each piece — because “we all know these days we don’t write code like that.”

FAQ

What’s the difference between ACP and UCP? Competing attempts at the same thing: standardizing commerce primitives so agents can complete a purchase. ACP is OpenAI’s, UCP is Google’s. Prio ran the same checkout under both with different schemas, and listed convergence as an open question.

Can an AI agent actually pay for something today? Only under tight constraints. Prio said ChatGPT payments go through a shared payment token and Gemini’s UCP path accepts Google Pay, with no support yet for more autonomous schemes like X402. AP2 adds a scoped mandate — max amount, currency, revocation URL, consent proof — but he listed its real usage as still forming.

Why push a product feed instead of exposing a search API? Neither ACP nor UCP supports a search-catalog call. Prio’s technical reason: with M merchants and N products, live search fans out to too many calls, while a pre-sent feed can be indexed in advance. Commercial reasons too — sponsored products, retail media, ranking.