{
  "video": {
    "id": "G7cgLjZtmMU",
    "title": "The Agentic Commerce Stack — Ahnaf Prio, Best Buy",
    "duration": 1238,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "G7cgLjZtmMU",
    "title": "The Agentic Commerce Stack — Ahnaf Prio, Best Buy",
    "one_liner": "A Best Buy engineering manager maps the acronym soup of agentic commerce — MCP, A2A, ACP, UCP, AP2 — onto what each actually does in a checkout flow, and demos a working customer-agent/merchant-agent stack built on those primitives.",
    "summary": "Prio argues that browser-driving shopping agents (screenshotting, reading the DOM, filling forms) were clunky, slow, brittle and set off merchants' fraud alarms, and that what actually works today is merchants exposing structured primitives — product feeds and checkout APIs — that agents call directly with no browser. He walks through the mental model: MCP for tool access, A2A for agent-to-agent messaging, ACP (OpenAI) and UCP (Google) as the competing commerce primitives, and AP2 as Google's agentic payment mandate spec. He demos 'Ginny', his orange tabby reimagined as a bakery merchant agent, running on Cerebras at 3,000 tokens/sec, showing the A2A calls, the MCP product_search tool call, the UCP checkout state machine and an AP2 token side by side with the ACP equivalent. The closing argument is that conversational commerce without evals is whack-a-mole, illustrated by his own demo refusing to invent a discount code.",
    "key_points": [
      "About 45% of all agent sessions on major providers like ChatGPT and Google Gemini are shopping-related; agentic shopping is a ~$7B industry today, projected to reach $65B by 2030.",
      "Browser-automation shopping agents (Claude Chrome extension, Atlas) 'kind of just didn't work' — clunky, slow, brittle, and an AI impersonating a browser 'is just firing up all the alarm bells' for merchant engineering teams, often getting stuck at the payment flow.",
      "The acronym stack decoded: MCP = how the agent discovers and calls merchant capabilities (products, product detail, loyalty); A2A = how customer agent and merchant agent (or domain-level agents) talk; ACP (OpenAI) and UCP (Google) = the commerce primitives/schemas; AP2 = Google's agentic payment protocol with scoped payment mandates.",
      "Neither ACP nor UCP currently supports a live search-catalog call — merchants must push a product feed ahead of time so it can be indexed. Prio's stated reasons: sponsored products, retail media and ranking, plus the M-merchants × N-products call explosion. Meta has a third, similar-but-different feed spec.",
      "Payments are deliberately not autonomous yet: ChatGPT accepts only a shared/delegated payment token, Gemini UCP only Google Pay; nothing supports X402-style autonomous payment because liability sits with the merchant and its payment processor.",
      "Live demo of 'Ginny' the bakery merchant agent (model served by Cerebras at 3,000 tokens/sec) showing A2A messages, an MCP product_search tool call, the UCP checkout state machine (not ready for payment → ready for payment → completed), an AP2 mandate with max amount, currency, revocation URL and single-use flag, the ACP equivalent side by side, a timeline view, and a feed comparison across UCP and Meta with catalog sync every few seconds.",
      "Merchant nuance the specs exist to capture: 'to us merchants, that's a second line item, buddy' — adding a second quantity is not the same as adding an item.",
      "Eval categories he recommends: behaviour evals (e.g. don't hand out discount codes or reveal who else is checking out a product), protocol compliance (feeds must conform or ChatGPT/Gemini won't accept them), latency benchmarks, and LLM-as-quality-judge.",
      "Stable today: MCP, A2A, and ACP/UCP being out there. Still forming: real AP2 usage, ACP/UCP convergence, identity and consent standards, and multi-agent checkout delegation."
    ],
    "takeaways": [
      "Stop building browser-driving shopping automation; expose (or consume) structured merchant primitives — product feed plus checkout API — so the agent never touches a cart UI.",
      "Push a product feed to the platforms rather than waiting for them to crawl your PDP or call a search catalog, and keep it in sync as inventory and attributes change; budget for three schemas (ACP, UCP, Meta) and a sync layer that translates one internal catalog into all of them.",
      "Even if you're only building your own on-site merchant agent, adopt the standardized primitives — they're well thought out and they let you sell externally on ChatGPT and Gemini later with the same plumbing.",
      "Write evals before shipping a conversational commerce agent: behaviour (discount codes, sensitive data leakage, off-topic use like the Chipotle agent answering programming questions), protocol compliance for the feeds, latency benchmarks, and LLM-as-judge on your best use cases.",
      "For payments, plan around delegated tokens / Google Pay now, and design your mandate structure (authorizer, allowed purchases, max amount, revocation URL, consent proof) the AP2 way so autonomous checkout is a swap rather than a rewrite."
    ],
    "topics": [
      "agentic-commerce",
      "mcp",
      "a2a",
      "protocols",
      "payments",
      "evals",
      "retail",
      "agents"
    ],
    "tools": [
      "Best Buy",
      "ChatGPT",
      "Google Gemini",
      "Google AI mode",
      "MCP (Model Context Protocol)",
      "A2A",
      "ACP (Agentic Commerce Protocol)",
      "UCP (Universal Commerce Protocol)",
      "AP2",
      "x402",
      "Google Pay",
      "Cerebras",
      "Claude Chrome extension",
      "Atlas",
      "Meta / Instagram / Facebook product feed",
      "Microsoft Copilot",
      "GoPuff",
      "Grok",
      "Ray-Ban",
      "Chipotle",
      "GitHub"
    ],
    "quotes": [
      {
        "text": "Right now about 45% of all agent sessions that happen within major providers like chat.gbt.com and Google Gemini are related to shopping.",
        "at": "01:43",
        "url": "https://www.youtube.com/watch?v=G7cgLjZtmMU&t=103s"
      },
      {
        "text": "If you are a merchant who's trying to sell stuff, any engineering department of that merchant will tell you an AI impersonating your browser is just firing up all the alarm bells.",
        "at": "03:31",
        "url": "https://www.youtube.com/watch?v=G7cgLjZtmMU&t=211s"
      },
      {
        "text": "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. But to us merchants, that's a second line item, buddy.",
        "at": "05:04",
        "url": "https://www.youtube.com/watch?v=G7cgLjZtmMU&t=304s"
      },
      {
        "text": "We've realized working with AI and conversational experiences without evals is playing whack-a-mole.",
        "at": "16:20",
        "url": "https://www.youtube.com/watch?v=G7cgLjZtmMU&t=980s"
      }
    ],
    "words": 3877
  },
  "summary_url": "/#G7cgLjZtmMU",
  "transcript": {
    "html": "/transcripts/G7cgLjZtmMU.html",
    "txt": "/transcripts/G7cgLjZtmMU.txt",
    "vtt": "/transcripts/G7cgLjZtmMU.vtt"
  }
}