{
  "video": {
    "id": "YIVkERhy8xo",
    "title": "We're the bottleneck, but we don't have to be — Ido Salomon, AgentCraft",
    "duration": 890,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "YIVkERhy8xo",
    "title": "We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps",
    "one_liner": "Humans supervising many coding agents are the real bottleneck. Ido Salomon argues that RTS-game skills (visibility, jump-to-attention, delegation, co-op play) are the fix, and he built AgentCraft, a game-style agent orchestrator, to show it.",
    "summary": "Agents are highly capable, but each one has to be steered, directed and reviewed, so at scale the human burns out and becomes the bottleneck. Salomon argues that the skills needed already exist in video games like Warcraft, Sims and StarCraft, where you supervise many units at once. He demos AgentCraft, which raises the ceiling in three steps: visibility, then autonomy, then collaboration. He also previews an experimental, mobile-game-like tool (working name 'loopers') meant to 'lower the floor' for the roughly 90% of people who don't use agents yet.",
    "key_points": [
      "The problem isn't agent capability. Running something like 25 Claude Code agents means steering, directing and reviewing each one, which is exhausting at scale: 'we are the bottleneck.'",
      "AgentCraft is a game-inspired orchestrator. Agents (Claude Code, OpenClaw, Codex, OpenCode) show up as units on a map. It can detect existing agents on your device or spawn new ones, and you prompt them from a side panel that also takes voice. Buildings on the map stand for features such as plugin/skill management, an integrated terminal and integrated Git.",
      "Visibility: a side panel shows each agent's task, what it last did and what it's doing now. The file system can be projected onto the map, with files shown as runes, so you can see which agent touched which file and when. That data also produces heat maps. As in RTS games, pressing spacebar jumps to the nearest thing that needs your attention, such as a question to answer or a plan to approve.",
      "Visibility alone is still exhausting. Two problems remain: there's too much to hold in your head, and babysitting 20 parallel tasks is draining. His answer is autonomy. Agents scan the codebase and propose tasks you accept as 'quests'. A high-level goal spawns an orchestrator that breaks the work down and runs it in isolation in a local container (what's now called 'software factories'). Loops run in the background, for example scanning Twitter or your GitHub to create features.",
      "Review kit: see the full change set file by file, or review visual evidence (videos and photos of what changed). You can also run several instances in parallel and pick the best implementation.",
      "Collaboration: you can host rooms ('war halls' or 'war rooms'). They run locally but can be shared through tunnels. Teammates, in his demo his wife acting as product designer, appear on the same map. You can hand off or fork their work, so you're not tied to Git. A shared notice board and chat connect both people and agents. You can also access it from mobile or Telegram.",
      "Unexpected feedback: kids and non-productivity people started orchestrating agents with it, including someone who 'flunked out of college because he played Starcraft'.",
      "Lowering the floor: an experimental tool (working name 'loopers') aims for a mobile-game level of complexity instead of file-level detail. It's built around projects you keep returning to and loops that rely on more autonomous agents. The basic flow is 'I prompt, I get the result', and the visualization is meant to be customizable because different people find different interfaces easy."
    ],
    "takeaways": [
      "Treat supervising agents as an attention-management problem. Build or adopt visibility (per-agent status, which files agents touch) and a single 'jump to what needs me' action instead of watching every tool call.",
      "Hand off breakdown and babysitting: let an orchestrator split a high-level goal into tasks running in isolated containers, and use background loops for proactive work.",
      "Speed up review with visual evidence (videos or screenshots of changes) and by running several implementations in parallel and picking the best.",
      "Share agent work with teammates who may be better placed to review it, through shared workspaces where people and agents can see each other's work and hand off or fork it.",
      "Design agent tools for less technical users too ('lower the floor'), not just for power users."
    ],
    "topics": [
      "agents",
      "agent orchestration",
      "human-in-the-loop",
      "developer experience",
      "gamification",
      "code review",
      "collaboration",
      "autonomous agents"
    ],
    "tools": [
      "AgentCraft",
      "MCP-UI",
      "MCP Apps",
      "Claude Code",
      "OpenClaw",
      "Codex",
      "OpenCode",
      "Git",
      "Telegram",
      "npx",
      "Twitter",
      "GitHub",
      "Warcraft",
      "StarCraft",
      "Sims",
      "Civilization"
    ],
    "quotes": [
      {
        "text": "in reality we are the bottleneck",
        "at": "01:34",
        "url": "https://www.youtube.com/watch?v=YIVkERhy8xo&t=94s"
      },
      {
        "text": "I took this concept uh too far and I created agent uh which is a game inspired orchestrator",
        "at": "02:14",
        "url": "https://www.youtube.com/watch?v=YIVkERhy8xo&t=134s"
      },
      {
        "text": "we also need to uh lower the floor",
        "at": "11:30",
        "url": "https://www.youtube.com/watch?v=YIVkERhy8xo&t=690s"
      },
      {
        "text": "So we are the bottleneck, but we don't have to be.",
        "at": "13:32",
        "url": "https://www.youtube.com/watch?v=YIVkERhy8xo&t=812s"
      }
    ],
    "words": 2894
  },
  "summary_url": "/#YIVkERhy8xo",
  "transcript": {
    "html": "/transcripts/YIVkERhy8xo.html",
    "txt": "/transcripts/YIVkERhy8xo.txt",
    "vtt": "/transcripts/YIVkERhy8xo.vtt"
  }
}