{
  "video": {
    "id": "UOcHfR3_tys",
    "title": "Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band",
    "duration": 1054,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "UOcHfR3_tys",
    "title": "Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band",
    "one_liner": "Vlad Luzin, CTO of Band, argues that a multi-agent system with remote agents is really a distributed system of non-deterministic microservices. MCP, A2A and Slack/Telegram bots leave each agent 'in digital solitary confinement', so agents need a shared collaboration layer built on conversations, with rooms, channels and participants, plus governance.",
    "summary": "Luzin argues that multi-agent coordination is a problem developers already have today, not a hypothetical one: anyone running one coding session to do the work and another to review it is acting as a 'Cisco router' between agents that can't talk to each other. He walks through the current options. Loop-engineering libraries, MCP (agents as stateless tools), A2A (client-server, with no discovery in the protocol) and messaging-platform bots all leave either plumbing work or agents that can only talk to one human. He says the fix needs a transport layer, continuity, runtime ID binding across frameworks, conversation-level abstractions and governance. He then demos Band, which provides those pieces, and Jam, a desktop app the company uses internally to develop its own software.",
    "key_points": [
      "Thesis: the future belongs to AI-to-AI communication within a business, between businesses, and between consumers and businesses. The agents will be built in different frameworks, written in different languages and deployed in different environments, and they will talk to each other with no human involved.",
      "Developers running parallel Claude/Codex sessions (one doing the work, one reviewing) are acting as a router between stateful agents that can't communicate. So multi-agent coordination is a problem of today, not a hypothetical one.",
      "Current protocols fall short. MCP calls agents as tools and is completely stateless, so sticky sessions are hard. A2A is client-server, so two agents that both send tasks must each be a client and a server. Chaining agents brings REST API timeouts, you need persistent queues to track messages, and discovery isn't part of A2A at all.",
      "Connecting an agent to a messaging platform is manual work from the docs: 5 steps for Telegram, 7 for Discord, 8 for Slack and 11 for WhatsApp. All you get is an agent that can talk to one person, usually you.",
      "A multi-agent system where every agent is remote is 'a distributed system of microservices where each microservice is non-deterministic'.",
      "What's needed: a transport layer (ordered, real-time delivery, retries); continuity and consistency hydration when pods or containers crash; runtime binding that maps thread, conversation and execution IDs across frameworks; abstractions raised from IP/URL/pub-sub to rooms, channels and participants, with deterministic routing within and across channels; and a governance layer (identity, audit).",
      "Band demo: Codex and Land Rush agents register and onboard programmatically, and an agent card appears. Codex sends a contact request to a personal assistant (Andy) belonging to a different user in a different registry, which needs bilateral consent. Once approved, Codex invites Andy into a conversation and exchanges messages.",
      "Jam demo: it captures the tasks Claude and Codex create for themselves, shows a live map of which architecture components agents are touching, and pings a human when one is needed. It also shows cost attribution, e.g. a full-stack developer local Claude session at $2,000 in tokens and an architect local Codex session at $600. Engineering-manager, developer and architect Claude instances review PRDs and SRS documents with no hand-coded loops, because models already understand how to communicate through messaging platforms."
    ],
    "takeaways": [
      "Count the copy-pasting between agent sessions as a real coordination cost you have today, not a future concern to defer.",
      "Before you rely on MCP or A2A for agent-to-agent work, check whether you need statefulness, two-way task sending, timeout handling, persistent queues and discovery. Neither protocol gives you all of these.",
      "Design multi-agent systems as distributed systems: plan for ordered delivery, retries, crash recovery and mapping IDs across frameworks.",
      "Have agents talk through conversation primitives (rooms, channels, participants) instead of hand-coding orchestration loops in Python. Models already know how to work in messaging-style conversations.",
      "Instrument agent traffic so you can attribute token cost and how much a human was actually involved, per developer and per team."
    ],
    "topics": [
      "multi-agent systems",
      "agent communication",
      "mcp",
      "a2a",
      "distributed systems",
      "agent governance",
      "developer tooling",
      "cost attribution"
    ],
    "tools": [
      "Band",
      "Jam",
      "MCP",
      "A2A",
      "ACP",
      "Claude",
      "Codex",
      "Anthropic",
      "Slack",
      "Telegram",
      "Discord",
      "WhatsApp",
      "Salesforce",
      "Databricks",
      "SAP",
      "Lambda",
      "Land Rush",
      "Cisco",
      "Python",
      "TypeScript"
    ],
    "quotes": [
      {
        "text": "They are in digital solitary confinement.",
        "at": "06:21",
        "url": "https://www.youtube.com/watch?v=UOcHfR3_tys&t=381s"
      },
      {
        "text": "multi-agent system where every agent is remote is basically a distributed system of microservices where each microservice is non-deterministic.",
        "at": "07:51",
        "url": "https://www.youtube.com/watch?v=UOcHfR3_tys&t=471s"
      },
      {
        "text": "You are basically acting as a router, a Cisco router or a switch between two stateful agents that do work on your behalf",
        "at": "03:23",
        "url": "https://www.youtube.com/watch?v=UOcHfR3_tys&t=203s"
      },
      {
        "text": "MCP, calling agents as tools, it's completely stateless. If you want two agents to be stateful, sticky sessions, good luck to you.",
        "at": "04:38",
        "url": "https://www.youtube.com/watch?v=UOcHfR3_tys&t=278s"
      }
    ],
    "words": 2989
  },
  "summary_url": "/#UOcHfR3_tys",
  "transcript": {
    "html": "/transcripts/UOcHfR3_tys.html",
    "txt": "/transcripts/UOcHfR3_tys.txt",
    "vtt": "/transcripts/UOcHfR3_tys.vtt"
  }
}