{
  "video": {
    "id": "48YUYDjwfYY",
    "title": "Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI",
    "duration": 1088,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "48YUYDjwfYY",
    "title": "Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI",
    "one_liner": "Rather than rationing tokens, let an automated harness pick the model for each task based on cost per task at equal quality. Cast AI says doing this with its Kimchi coding agent saved 2.5x over three months, even though token use went up.",
    "summary": "Cast AI's Laura (co-founder and president) and Kimchi's head of engineering argue that capping developers' tokens is the wrong answer to exploding LLM bills. A manager's job is to make tokens effectively unlimited and cheap, and the way to do that is a harness that routes each task to whichever model gives the lowest cost per task, which is not the same as the lowest price per token. They show results from three months of internal use: the harness moved on its own from Kimi 2.6 to MiniMax 3 within a few weeks of June. They then demo the rest of the Kimchi platform: Ferment for long autonomous tasks scored on output quality, Teleport for remote sandboxed sessions, and Studio for team kanban-style collaboration.",
    "key_points": [
      "Examples of runaway token spend from the news: a company in India spent $500M on Anthropic alone in one month, and Uber's CTO said they used a full year's Anthropic budget in four months.",
      "Price per token is misleading; what matters is cost per task at the same quality, citing a university paper. Gemini 3 Flash costs about $3.5 per million tokens (blended) but $75 per task. MiniMax 2.7 is cited at 1.5 per token but $148 per task.",
      "Over three months at Cast AI (300 employees, two-thirds of them developers), Kimchi delivered 2.5x savings against what the cloud bill would have been, for the same outcome. Token volume grew by 1.5x over the same period.",
      "The harness changes models by itself. On June 3, Kimi 2.6 was its most-picked model. A shift happened on June 12, and by June 21 MiniMax 3 was the top pick. The speakers argue humans could not evaluate and adopt new models that fast.",
      "Ferment is an open-source construct for long tasks, run autonomously for more than two hours on the assumption that a human checks in only every 2-3 hours. It asks questions up front, splits the work into milestones, then builds, checks for breakage and fixes in a loop. A higher-order model scores the output, and an artifact is complete only at a B or better; you can ask it to push for an A.",
      "Deployment deliberately stops at staging so a human stays in the loop. The planned next step is automatic production deploys gated on observability, such as Kubernetes pods not crashing and health checks passing.",
      "Kimchi Teleport syncs your local environment to a container in a Kubernetes cluster on Google Cloud, so sessions keep running when the laptop is closed. It works from laptop or mobile and can be shipped on-prem, for example for fintech. It started when someone's laptop battery died mid-flight. 62% of Cast AI engineers now code only through Teleport.",
      "Kimchi Studio runs on Teleport sandboxes and shows a team's agent sessions on a browser kanban board (backlog, in progress, review). Teammates such as a peer engineer or product manager can review plans and pick up sessions waiting for review. It is aimed at pizza teams of 5-10."
    ],
    "takeaways": [
      "Compare models on cost per task at equal quality, not on price per token.",
      "Don't cap developers' tokens; route each task to the cheapest model that meets a quality bar, and let the routing update itself as new models come out.",
      "Gate agent output on an automated quality score (for example, B or better) and loop until it passes, rather than having humans read huge AI-generated diffs.",
      "Run long agent sessions in remote sandboxes so they survive a closed laptop or a dead battery, and make plans and specs reviewable by the team before work goes too far.",
      "For now, keep a human gate before production. Automate the move to production only once observability checks such as pod health and health checks can verify deploys."
    ],
    "topics": [
      "coding agents",
      "model routing",
      "llm cost optimization",
      "agent harness",
      "remote sandboxes",
      "agentic engineering",
      "code quality scoring",
      "team collaboration"
    ],
    "tools": [
      "Kimchi",
      "Cast AI",
      "Kimchi Ferment",
      "Kimchi Teleport",
      "Kimchi Studio",
      "Anthropic",
      "Claude Code",
      "Gemini 3 Flash",
      "MiniMax 2.7",
      "MiniMax 3",
      "Kimi 2.6",
      "Google Cloud",
      "Kubernetes",
      "Uber"
    ],
    "quotes": [
      {
        "text": "Our job is not to prevent the developer to use coding agent.",
        "at": "02:18",
        "url": "https://www.youtube.com/watch?v=48YUYDjwfYY&t=138s"
      },
      {
        "text": "what we believe is that reading code is not enough anymore.",
        "at": "11:01",
        "url": "https://www.youtube.com/watch?v=48YUYDjwfYY&t=661s"
      },
      {
        "text": "I am puzzled that sometimes innovations are always the simplest one.",
        "at": "12:57",
        "url": "https://www.youtube.com/watch?v=48YUYDjwfYY&t=777s"
      },
      {
        "text": "62% of our engineers are only using teleport when they are coding now for that problem.",
        "at": "14:05",
        "url": "https://www.youtube.com/watch?v=48YUYDjwfYY&t=845s"
      }
    ],
    "words": 3330
  },
  "summary_url": "/#48YUYDjwfYY",
  "transcript": {
    "html": "/transcripts/48YUYDjwfYY.html",
    "txt": "/transcripts/48YUYDjwfYY.txt",
    "vtt": "/transcripts/48YUYDjwfYY.vtt"
  }
}