{
  "video": {
    "id": "YZQsWVeN3rE",
    "title": "Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers — Alex Bauer, Upside.tech",
    "duration": 1029,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "YZQsWVeN3rE",
    "title": "Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers — Alex Bauer, Upside.tech",
    "one_liner": "Upside's Alex Bauer reframes AI hallucination as a trust problem and argues you fix it by managing agents like people — scaffold them with written anchor docs, make them consult a \"librarian\" for your company's real definitions before answering, and settle no-right-answer questions with a jury of independent analyst agents weighed by a consensus judge.",
    "summary": "Bauer opens with a bedtime-story allegory of a go-to-market team fighting security review, conflicting metric definitions and finally the question \"how is the business doing?\", then unpacks three real solutions his company ships or uses. His thesis is that the hallucination problem grew up into a trust problem — Claude asked about revenue never says \"I'm not sure,\" it hands you a wrong answer that looks exactly like a right one — and that the practical fixes are the ones we already use with human colleagues: state intent, document how the world works, give people a knowledgeable guide, and get a second opinion before believing anything. He walks through scaffolding a website rebuild with maintained \"anchor assets\", a librarian that injects just-in-time company knowledge before an agent queries data, and a jury-and-judge workflow for multi-touch attribution. A bonus section on \"agent tiers\" argues that AI crowbarred into an existing subscription price has no margin for a real reasoning model.",
    "key_points": [
      "The hallucination problem has an \"older sibling\": a trust problem. Ask Claude to report on revenue and it doesn't hedge — it returns a confident wrong answer that is indistinguishable from a right one.",
      "One practical tip above all others: use commander's intent when you prompt — armed-forces doctrine of telling agents *why* you want something. But agents trained on human material micromanage themselves, so \"don't tell Claude to improve itself\" — pull it back to the why.",
      "Rebuilding the company website in YOLO mode (\"here are the sources, go read everything, please give me my website\") failed even using Claude's plan mode first. The fix was scaffolding: a maintained set of \"anchor assets\", notably a product capabilities reference where each card states what a capability does, why it matters per defined persona, and carries citations across every connected system so he can verify nothing important was hallucinated. At his previous 600-person company the same job meant several product marketers, a designer, consulting PMs and a web team over 2 months.",
      "He also keeps a \"Persona Bench\" — a set of agents inhabiting each important persona that can review work on demand and give their opinions — plus generic AI-slop scales, of which \"there are a ton of these floating around\".",
      "The librarian: instead of letting an agent answer \"how much pipeline did we create in Q1?\" by assuming Q1 means January–March on created date, the agent consults a librarian first, which holds documentation, a library of company knowledge items, the schema, and prior failed queries. That gives just-in-time memory — this company's fiscal year is February through April, and pipeline means stage two or later — and returns a trustworthy answer with citations.",
      "Jury-and-judge workflow for multi-touch attribution, a class of GTM problem with no empirically correct answer: the agent spins up a team of independent analysts who each examine the data alone and produce an evidence-cited opinion; a consensus judge treats those as input rather than fact, weighs the reasoning quality of each analyst, and expands the jury if consensus is insufficient. The judge does no research of its own.",
      "Multi-touch attribution took the company 2 years to get close to; along the way they built an AI-native data layer that turned out to be what customers needed for everything else, and returning to attribution \"enabled by Opus\" finally produced impressive results.",
      "Agent tiers: \"you can't fix stupid.\" Slackbot's newly released MCP client looked like a way to give everyone with Slack access to the Upside librarian but proved \"horrifically stupid\" — the general case being AI crowbarred into a pre-existing subscription price, where the margin leaves no room for an intelligent reasoning model. Tier two means a powerful model with sub-agents, plan mode, full MCP support and file editing.",
      "Go-to-market teams historically had ideas but a toolbox of spreadsheets and PowerPoint; AI gives the non-engineers \"an infinite supply of valedictorian interns with computer science degrees\", and Claude does for building what the bicycle did for mobility."
    ],
    "takeaways": [
      "Put the why in your prompts — commander's intent — and when the agent starts micromanaging itself or \"improving itself\", pull it back to the intent.",
      "Define the structure before you turn the model loose: maintain anchor assets (personas, a cited product capabilities reference) as durable documentation rather than YOLO-ing a large build from raw sources.",
      "Insert a librarian lookup ahead of any data question so company-specific definitions — fiscal calendar, what counts as pipeline, prior failed queries — arrive as just-in-time memory, and return answers with citations you can follow back.",
      "For questions with no empirically correct answer, don't let one agent answer: run independent analysts with evidence-cited opinions into a consensus judge that weighs reasoning quality and expands the jury when there isn't enough agreement.",
      "Audit the tier of harness and model your team uses for important work — don't let them run on the ChatGPT web interface or on AI bolted into a flat-priced SaaS subscription and expect a good result."
    ],
    "topics": [
      "agents",
      "trust",
      "go-to-market",
      "prompting",
      "multi-agent",
      "context-engineering",
      "attribution",
      "mcp"
    ],
    "tools": [
      "Claude",
      "Claude plan mode",
      "Opus",
      "Upside",
      "Slack / Slackbot MCP client",
      "MCP",
      "ChatGPT",
      "Persona Bench",
      "xAI"
    ],
    "quotes": [
      {
        "text": "It says, \"Here you go.\" And it gives you a wrong answer that looks exactly like being right.",
        "at": "07:53",
        "url": "https://www.youtube.com/watch?v=YZQsWVeN3rE&t=473s"
      },
      {
        "text": "the main thesis of the talk today is actually, when in doubt, manage your agents like other humans",
        "at": "08:37",
        "url": "https://www.youtube.com/watch?v=YZQsWVeN3rE&t=517s"
      },
      {
        "text": "define the structure first, and then turn Claude loose. Don't try and YOLO it from the beginning.",
        "at": "11:25",
        "url": "https://www.youtube.com/watch?v=YZQsWVeN3rE&t=685s"
      },
      {
        "text": "In my experience, you can't fix stupid.",
        "at": "15:27",
        "url": "https://www.youtube.com/watch?v=YZQsWVeN3rE&t=927s"
      }
    ],
    "words": 3360
  },
  "summary_url": "/#YZQsWVeN3rE",
  "transcript": {
    "html": "/transcripts/YZQsWVeN3rE.html",
    "txt": "/transcripts/YZQsWVeN3rE.txt",
    "vtt": "/transcripts/YZQsWVeN3rE.vtt"
  }
}