{
  "video": {
    "id": "MSbacZ99E14",
    "title": "High Throughput Agentic Engineering with ex-Meta L8 Principal",
    "duration": 3237,
    "upload_date": "2026-09-13",
    "channel": "Kun Chen",
    "source": "https://youtu.be/MSbacZ99E14?is=msmvz0KIwijCYUhN",
    "event_format": ""
  },
  "analysis": {
    "video_id": "MSbacZ99E14",
    "title": "High Throughput Agentic Engineering with ex-Meta L8 Principal",
    "one_liner": "An ex-Meta L8 principal engineer demos his \"first mate\" pattern: he talks to exactly one agent session, which dispatches and supervises a fleet of crew mates across 36 projects, routing each task to a harness/model by rule and remaining quota.",
    "summary": "Kun walks through a live working session in his open-source First Mate workflow, where a single always-available orchestrator agent spawns \"crew mates\" for every task so he never context-switches between sessions. He shows the dispatch rule file that picks a harness and model per task type, quota-axi picking whichever model has the most subscription quota left, Lavish interactive HTML artifacts for reviewing visual prototypes, and per-project merge policies ranging from \"direct PR + yolo\" to the adversarial \"no mistakes\" validation pipeline. The through-line argument is that human attention and time — not tokens or context — are the real bottleneck, so you should stop hand-optimizing context windows and spend that bandwidth deciding what to build.",
    "key_points": [
      "The core pattern: one agent session (\"first mate\") holds all intent and context across 36 projects and orchestrates every other agent; he never juggles sessions. When first mate gets too busy orchestrating to talk to him, he creates \"second mates\" that own a domain (ship app, Eddie's Wallet, axi) — each second mate is itself a full first mate with its own crew mates, memory and instructions.",
      "A custom Pi extension called `calm` hides all tool calls and shows only a little floating boat, because watching tool calls is \"distractions and noise\" — `/calm` toggles detail back on when he actually wants it.",
      "Dispatch is rule-driven from `config/crew-dispatch.json` (written by first mate from his stated preferences): new features on his paid iOS app → Fable; tasks needing image generation → Codex + GPT 5.6 (Codex has a native image tool); genuinely hard architecture/planning → Fable, Kimi K3 or Astra; simple bug fixes with known root cause → Luna, Sonnet, Cursor+Grok 4.6; default → Claude Opus or Cursor+Grok 4.6 high.",
      "When a rule lists multiple models, first mate calls quota-axi (his own tool, TUI for humans, CLI for agents) and picks whichever subscription has the most quota left — \"with people like Tibo constantly resetting codex quota it's really hard to keep track of that.\"",
      "Prototypes come back as Lavish interactive HTML boards — screenshots of current vs. variants, side-by-side compare, explanations, and a decisions list at the bottom. He asked for 3 variants each of snowy mountains, northern lights and bird types for his Fly With Me game; feedback typed into Lavish goes straight back to the crew mate, bypassing first mate (which then thinks the decision is still open).",
      "Per-project merge policy lives in `data/projects.md`: toy projects get \"direct PR + yolo\" (agent judges and self-merges), higher-stakes ones require the \"no mistakes\" adversarial validation pipeline. His heuristic: if you wouldn't have asked a human to review this change, you don't need no-mistakes — it takes a long time and a lot of tokens.",
      "A no-mistakes PR reads as intent, what changed, a risk assessment, and live end-to-end validation scenarios with evidence (the demo PR ran 9 of 10, skipping a documentation-only one). Low risk + well tested means he ships without reading a line of the diff.",
      "First mate is summonable from Discord and X — @-mentioning the bot relays into the same local session, which replies in-thread (\"Aye captain, looking into the treehouse pre-destroy hook\") and dispatches the investigation. It also silently retried a PR creation that failed on a GitHub 500 without involving him.",
      "Two built-in catch-up skills: `bearings` (snapshot of everything underway across the fleet plus pending \"captain's calls\") and `ahoy` (cheap summary of everything first mate said since your last message, plus open decisions) — his fix for chat UX losing things while you're off reviewing an artifact.",
      "Model opinion: he prefers Grok 4.5 over the newer 4.6 — 4.5 is \"faster and more efficient and more straight to the point\", while 4.6 is \"a little bit like Opus 5\" with a \"weird personality where it doesn't quite speak like a human.\""
    ],
    "takeaways": [
      "Stop hand-managing context windows. Set one threshold and forget it — he sets `CLAUDE_CODE_AUTO_COMPACT_WINDOW` to 500k (Claude Code's 1M default is \"a little bit too high\"; Grok's 500k is fine as-is) and spends the recovered mental bandwidth on deciding what to build.",
      "Put model selection in a config file, not in your head: write per-task-type rules mapping to harness+model, list several acceptable models per rule, and let the orchestrator pick by remaining quota so no subscription goes to waste.",
      "Tier your guardrails per project rather than globally. Ask \"would I have asked a human to review this?\" — if no, let the agent PR-and-merge; reserve the expensive adversarial validation pipeline for changes that ship straight to users.",
      "Have agents deliver prototypes as reviewable visual artifacts with variants and an explicit decisions list, so you validate by looking and picking rather than by reading code or diffs.",
      "Don't assume a newer model is better — use it for a while, form your own sense of how it behaves, and read what others have found before switching defaults.",
      "Queue work instead of waiting: use follow-up prompts (alt+enter in Pi) to stack tasks behind the current one and steer (enter) only when you need to interrupt — the orchestrator stays available to talk because it isn't doing the work itself."
    ],
    "topics": [
      "agents",
      "agent-orchestration",
      "parallel-agents",
      "model-routing",
      "context-management",
      "code-review",
      "developer-workflow",
      "prototyping"
    ],
    "tools": [
      "First Mate",
      "Herder",
      "Pi",
      "Claude Code",
      "Codex",
      "Cursor",
      "Backpass",
      "quota-axi",
      "Lavish",
      "no mistakes",
      "Treehouse",
      "Fly With Me",
      "Eddie's Wallet",
      "Ship",
      "Opus 5",
      "Fable",
      "Sonnet",
      "Grok 4.5",
      "Grok 4.6",
      "GPT 5.6",
      "Kimi K3",
      "Astra",
      "Luna",
      "Discord",
      "X",
      "GitHub",
      "tmux",
      "Mac Mini",
      "Meta",
      "Microsoft",
      "Atlassian"
    ],
    "quotes": [
      {
        "text": "the core idea is that uh I only talk to one agent which is the first mate and the first mate will orchestrate all the other agents for me so that I don't have to juggle between all the sessions and context switch all the time",
        "at": "01:05",
        "url": "https://www.youtube.com/watch?v=MSbacZ99E14&t=65s"
      },
      {
        "text": "our most precious resource right now, the biggest bottleneck across everything we do with agents is our attention and time",
        "at": "34:00",
        "url": "https://www.youtube.com/watch?v=MSbacZ99E14&t=2040s"
      },
      {
        "text": "If your brain is thinking about the context window all the time, you don't have enough time to think about what to build.",
        "at": "35:56",
        "url": "https://www.youtube.com/watch?v=MSbacZ99E14&t=2156s"
      },
      {
        "text": "A good way to think about this is like for this code change, would you have asked a another human to review this code? Uh if you would not have have another code reviewer, then you probably don't need no mistakes.",
        "at": "40:19",
        "url": "https://www.youtube.com/watch?v=MSbacZ99E14&t=2419s"
      }
    ],
    "words": 10957
  },
  "summary_url": "/#MSbacZ99E14",
  "transcript": {
    "html": "/transcripts/MSbacZ99E14.html",
    "txt": "/transcripts/MSbacZ99E14.txt",
    "vtt": "/transcripts/MSbacZ99E14.vtt"
  }
}