{
  "video": {
    "id": "toq-jyGLZDk",
    "title": "Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND",
    "duration": 1033,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "toq-jyGLZDk",
    "title": "Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND",
    "one_liner": "Connecting agents across machines is a distributed-systems problem, not a protocol problem: you need real-time ordered transport, persistence and hydration, runtime ID binding, a registry, and governance. BAND says its multi-peer interaction layer handles all of that so you don't have to.",
    "summary": "Vlad Luzin, co-founder and CTO of BAND, argues that the future is AI-to-AI communication within businesses, between businesses, and between consumers and businesses, with autonomous, always-on agents spread across the globe. He says today's multi-agent practice keeps the agents isolated. With adversarial agents, you are the router moving packets between sessions. Loop engineering hands that job to a script. Messaging platforms only connect an agent to its owner, and protocols such as MCP and A2A are too low-level to build on at scale. He lists the distributed-systems problems that need solving, then demos BAND's interaction layer: Claude, Codex and Claude Code agents from two users onboard in seconds, discover each other, and collaborate in shared conversational spaces with a human in the loop.",
    "key_points": [
      "Adversarial agents (one session plans or writes code, another reviews) mean 'you are basically a Cisco router and a switch' moving packets between two stateful agents. Loop engineering is the same thing, except a piece of open-source Python or TypeScript prompts the agents back and forth for you.",
      "Multiple agents exist because of transformer limits, which he calls the 'single agent bottleneck': confirmation bias, attention dilution, context fragmentation and recall degradation. A 1M or 2M token context window won't fix performance.",
      "Messaging platforms aren't the answer. Connecting an agent takes 5 manual steps for Telegram, 7 for Discord, 8 for Slack and 11 for WhatsApp, and even then you only get an agent talking to you. These platforms are built for humans and actively block bot-to-bot connections.",
      "Protocols are too low-level. MCP calls are stateless, so you can't go back and ask the agent what happened. A2A is client-server and one-way, so two-way traffic needs both a client and a server on each side. Chained REST calls hit timeouts, and discovery and queues are missing (A2A is still in draft).",
      "What a real solution needs: real-time, ordered transport, because LLMs expect ordered messages; retries; persistence and hydration, because an agent is a microservice whose pod can crash; and runtime binding that maps thread, session, execution and run IDs across frameworks such as LangGraph, Codex, Claude and CrewAI.",
      "You have to work above IPs, ports, URLs and pub/sub topics, at the level of conversations, participants, channels or rooms, with message routing that treats agents as first-class citizens. Enterprises also need governance: identity, plus observability that covers the tool calls an agent made after receiving a message, not just the messages themselves.",
      "BAND's launch is pitched as a global interaction and collaboration layer for any agent, platform, language or environment. It is multi-peer rather than peer-to-peer, and it includes a registry, persistence, channels, platform-side message filtering, security and observability, with humans able to join.",
      "The recorded demo: agents onboard and their agent cards appear 'seconds later'. A contact request between two users' agents needs bilateral consent. A Codex agent invites Vlad's personal Claude assistant into a real-time conversational space. A Claude Code terminal session creates a space, invites Codex and a human, and the agents build and review a small website. The owner has full visibility of any conversation their agent joins."
    ],
    "takeaways": [
      "Recognise that running a planner and a reviewer session side by side is already a multi-agent system, and that you are currently the routing layer.",
      "Don't expect Slack, Discord, Telegram or WhatsApp integrations to give you agent-to-agent communication. They connect one agent to one human.",
      "Treat MCP and A2A as low-level building blocks. On their own they leave statefulness, two-way traffic, timeouts, discovery and queues for you to solve.",
      "When designing agent-to-agent communication, plan for ordered real-time delivery, crash recovery through persistence and hydration, cross-framework ID mapping, and observability down to tool calls.",
      "Keep a human in the loop: require consent before agents from different owners connect, and give owners full visibility into the conversations their agents take part in."
    ],
    "topics": [
      "multi-agent systems",
      "agent communication",
      "loop engineering",
      "distributed systems",
      "agent protocols",
      "observability",
      "human in the loop"
    ],
    "tools": [
      "BAND",
      "Claude",
      "Claude Code",
      "Claude SDK",
      "Codex",
      "Codex SDK",
      "Anthropic",
      "Slack",
      "Microsoft Teams",
      "Discord",
      "WhatsApp",
      "Telegram",
      "MCP",
      "ACP",
      "A2A",
      "Google",
      "LangGraph",
      "CrewAI",
      "Hermes"
    ],
    "quotes": [
      {
        "text": "you are basically a Cisco router and a switch moving packets between these two stateful agents.",
        "at": "02:40",
        "url": "https://www.youtube.com/watch?v=toq-jyGLZDk&t=160s"
      },
      {
        "text": "So, basically, your agents are still alone in a kind of digital solitary confinement.",
        "at": "05:18",
        "url": "https://www.youtube.com/watch?v=toq-jyGLZDk&t=318s"
      },
      {
        "text": "You're not building a multi-agent system. You're wasting your time on a boring boring stuff.",
        "at": "07:02",
        "url": "https://www.youtube.com/watch?v=toq-jyGLZDk&t=422s"
      },
      {
        "text": "it's loop engineering without 500,000 lines of a price book Python or Python script code",
        "at": "15:44",
        "url": "https://www.youtube.com/watch?v=toq-jyGLZDk&t=944s"
      }
    ],
    "words": 2884
  },
  "summary_url": "/#toq-jyGLZDk",
  "transcript": {
    "html": "/transcripts/toq-jyGLZDk.html",
    "txt": "/transcripts/toq-jyGLZDk.txt",
    "vtt": "/transcripts/toq-jyGLZDk.vtt"
  }
}