{
  "video": {
    "id": "REascnFlq_8",
    "title": "Agents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town",
    "duration": 1277,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "REascnFlq_8",
    "title": "Agents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town",
    "one_liner": "Town CTO Jean-Denis Greze argues that \"agent-to-agent\" is really a search problem — how closely a multi-agent system can approximate one omniscient agent with all the world's data in its context window — and walks through five strategies for getting cross-silo data into that final LLM call without violating privacy.",
    "summary": "Greze reframes multi-agent systems as context engineering: every LLM system is a search problem whose goal is having exactly the right information in the context window just before the final answer or tool call. The ideal is a single agent with access to all the world's information (a Coase-theorem world with no transaction costs), which is impossible only because of privacy and security — so the test of any multi-agent design is how well it approximates that. He grades five strategies (shared trust boundaries, privacy-preserving custom tools, shared silos with a \"sweeper AI\", humans as approval conduits, and a black-box agent that only asks permission at the disclosure step) against two questions: does it need fewer humans over time, and does it get better as models get better. His bet is on AI-maintained shared wikis now, and on an \"auto mode\" for privacy decisions that scales with model capability.",
    "key_points": [
      "Reframe: \"most LLM systems are just a search problem\" — the whole job is making sure the context window holds the right information right before a tool call or a response to the user. The progression: humans hand-populating context 4 years ago → RAG search tools → agentic search across many tools today.",
      "The benchmark for any multi-agent system is a thought experiment: one agent, one context window, access to every person's email and every company's and government's information. That is the ideal multi-agent world; the only thing blocking it is privacy and security, not context length. He invokes the Coase theorem — full information plus zero transaction costs yields the economically ideal outcome.",
      "Strategy 1, access within a trust boundary: he and his wife share an agent that reads both their inboxes (including pre-marriage email); at work, an HR agent with the access of the lowest-privileged HR employee. Popular with IT and security teams because it's the same model as SaaS security, but it fails both of his tests — it doesn't need fewer humans over time and doesn't improve as models improve. \"You've just created a new silo cuz a human thought about it.\"",
      "Strategy 2, custom tools that trade power against privacy: instead of exposing everyone's Gmail to answer \"is anyone at my company connected to the finance team at Acme Corp?\", build a tool that reads all inboxes but returns only a relationship-strength score per person for a given domain and role; the agent then Slacks the top match (Bob) to ask for an intro to Jane the CFO. Town ships tools like this, opt-out, chosen for natural network effects. Another example: letting colleagues drop draft emails into your inbox. Still manual — humans must invent the tool and explain the trade-off.",
      "Strategy 3, shared silos plus a \"sweeper AI\" — the one idea he singles out as working really well: an AI inside each private silo holds a policy of what must stay private plus descriptions of the shared spaces, and at end of day pushes new shareable information out to company-public spaces. Deciding what's shareable is either ask-a-human approval or LLM policy enforcement; he predicts LLM-enforced policies in production within 6 months at 10–50-person high-trust companies (with finance and HR data as the clear no-share categories), not at Fortune 500s.",
      "Strategy 4, humans as conduit (classic agent-to-agent) spams everyone: asking a 100-person company one question pings 100 people on Slack for approval. Strategy 5 fixes this with a black box — an LLM whose trace nobody can see searches all silos automatically, works out the answer or the pending tool call, then asks only the people whose information it actually needed. In the intro example, 20 people come back connected, it ranks them by email contacts, picks Bob, and only Bob gets asked.",
      "Failure modes he names: prompt injection planted in a more-open silo and pulled out by agentic search; shared wikis going off the rails from one LLM mistake that then poisons them forever (his own personal wiki still calls his agent \"Apex\" a month after he renamed it \"Ivy\"); wrong disclosures that get someone fired or get you sued; and the fact that a black box can't truly be a black box — someone in the CISO suite will need audit access.",
      "The frontier is \"auto\": coding went from approving everything, to YOLO, to Anthropic's auto mode, and privacy across silos will follow the same path — low-sensitivity information flows automatically, sensitive disclosures stay human-approved, and the auto zone widens as models improve. The open, unsolved prize is cross-company: he cites an unnamed firm getting investment banks to share private data on private companies for lending, with each side's agents deciding what the other can access."
    ],
    "takeaways": [
      "Judge any multi-agent or agent-to-agent design by two questions: does it require fewer humans over time, and does it get better as the models get better? Trust-boundary agents and hand-built privacy tools fail both — they're fine now, but they're not the end game.",
      "Build a sweeper AI rather than hoping people fill the wiki: give each private silo an agent with an explicit stay-private policy and a description of the shared spaces, and have it publish new shareable information into company-shared wikis, databases or repo skills on a daily cycle.",
      "Design a low-sensitivity zone where you are explicitly okay with the LLM deciding what to share, and get comfortable with it — a system built that way automatically scales with model capacity as the zone widens, instead of needing a rebuild.",
      "When an agent needs data from many people's silos, don't broadcast approval requests. Let a trace-private agent search everything first, then ask only the one or two people whose information the final tool call actually depends on.",
      "For any privacy-preserving tool, ask what the narrowest useful return value is — a relationship-strength score rather than the underlying emails — and make it opt-out, so it breaks silos without exposing them."
    ],
    "topics": [
      "agents",
      "multi-agent",
      "agent-to-agent",
      "context-engineering",
      "privacy",
      "agentic-search",
      "enterprise-ai",
      "network-effects"
    ],
    "tools": [
      "Town",
      "Plaid",
      "Dropbox",
      "Gmail",
      "Google",
      "LinkedIn",
      "Slack",
      "Airtable",
      "Anthropic"
    ],
    "quotes": [
      {
        "text": "most LLM systems are just a search problem.",
        "at": "01:22",
        "url": "https://www.youtube.com/watch?v=REascnFlq_8&t=82s"
      },
      {
        "text": "So, this actually if there's one good idea in this talk that I think works really well is this. It's a sweeper AI.",
        "at": "10:03",
        "url": "https://www.youtube.com/watch?v=REascnFlq_8&t=603s"
      },
      {
        "text": "I think in coding we used to approve everything. Then we were like, \"YOLO, live dangerously.\" And now the gods at Anthropic have granted us auto mode.",
        "at": "17:34",
        "url": "https://www.youtube.com/watch?v=REascnFlq_8&t=1054s"
      },
      {
        "text": "do I trust the future where agents make all the decisions about privacy? I don't know about that. I just think it's a it is for better or worse the direction things are going.",
        "at": "20:41",
        "url": "https://www.youtube.com/watch?v=REascnFlq_8&t=1241s"
      }
    ],
    "words": 4974
  },
  "summary_url": "/#REascnFlq_8",
  "transcript": {
    "html": "/transcripts/REascnFlq_8.html",
    "txt": "/transcripts/REascnFlq_8.txt",
    "vtt": "/transcripts/REascnFlq_8.vtt"
  }
}