{
  "video": {
    "id": "pMggiOb18tc",
    "title": "The Golden Age of AI Engineering — Alexander Embiricos & Romain Huet & Peter Steinberger, OpenAI",
    "duration": 1513,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "pMggiOb18tc",
    "title": "The Golden Age of AI Engineering — Alexander Embiricos & Romain Huet & Peter Steinberger, OpenAI",
    "one_liner": "OpenAI's Codex team argues engineers aren't being automated away but empowered — and Peter Steinberger shows the shift in practice: stop juggling 10 terminals and watching code fly by, and instead run a long-lived manager agent that delegates to workers while you work the outer loop.",
    "summary": "Romain Huet and Alexander Embiricos argue that because models now take on long, hard goals until they're done, the job returns to the roots of engineering — taste, judgement and imagination — and the product goal is explicitly \"not to automate engineers\" but to maximally empower them, via a chat-first surface plus a collaborative UI you drop into. They walk the Codex stack layer by layer (Responses API, open-source harness, AGENTS.md, app server, plugins) to make the point that OpenAI ships itself the same thing it ships developers, then cover \"value maxing\": cost, speed and where agents run. Peter Steinberger closes with the operational thesis: server-side compaction, delegation and triggers together form a loop, and once you have it your bottleneck moves from tokens to compute to attention — the one thing you can't buy more of.",
    "key_points": [
      "Model cadence went from a new model roughly every 15 months to roughly every 6 weeks; the 5.6 series preview launched the week before the talk.",
      "The capability ladder as Embiricos tells it: completion → inline prediction → Command-K edits → models testing their own work → models taking on long, hard goals until they're done. At Dev Day 2024 Huet's o1-preview drone demo couldn't run or verify its own code; by Dev Day 2025 models controlled a live camera and lighting system.",
      "The product bet is two modalities — chat (\"I think chat is underrated\") plus a hands-on collaborative UI — because a CLI can't host a collaborative interface and an IDE gets the order wrong by starting with code instead of conversation. Internal and external skeptics who said they'd \"never leave my terminal\" now use the Codex app.",
      "Codex is deliberately layered and open: the harness is open source and used in model post-training, AGENTS.md was named so other agents could adopt it, models aren't hardcoded so you can swap in an open model with the same loop, and app server — the same path OpenAI's VS Code extension and Codex app use — is open source. Compaction was built into the API first so developers got the same primitive.",
      "Cost/speed numbers: GPT 5.6 Terra delivers GPT 5.5-level intelligence at half the cost; Luna runs $1 per million input tokens and $6 per million output tokens; GPT 5.6 Soul on Cerebras hits 750 tokens/second — roughly a substantial PR in 10 seconds, enough to run five or six parallel approaches and pick the best.",
      "Steinberger's January setup was 10+ terminal windows where he was the scheduler, router and memory; now he talks to one long-running manager agent that spawns workers, and drops down to pair only for tricky work.",
      "Three things made the manager pattern work: server-side compaction (long tasks reliable enough to stop optimizing for the first session), coordination (one thread creating and steering projects), and automation that can wake the manager on an event — persistent context, delegation, triggers.",
      "Concrete loop demoed in narrative: an issue lands on an open-source repo, the manager reads it against the project's goals/notes/vision, spawns a worker to implement and test, another agent reviews, and it surfaces to the human as a PR plus diff, maybe a video or a running build to VNC into. Paul runs a pinned \"chief of staff\" that wakes every 10 minutes to coordinate GitHub work."
    ],
    "takeaways": [
      "Stop watching the agent generate code — that was necessary with earlier models so you could hit escape and steer, but current models understand intent well enough that supervising token-by-token is wasted attention.",
      "Move your default from pairing with one agent (or managing N terminals) to managing a single long-lived manager that delegates to workers, and reserve direct pairing for genuinely tricky work.",
      "Connect the agent to what happens before coding (why the work needs doing) and after it (review, deploy) — that's how agents begin more work and land more work, rather than just assisting mid-task.",
      "Offload the compute bottleneck before it becomes yours: run agent test suites on separate test boxes instead of melting your laptop, and stop leaving it open overnight so agents can keep running.",
      "Optimize for value maxing, not token maxing — pick the cheap/fast frontier tiers deliberately, and use raw speed to fan out five or six parallel attempts and pick the best rather than just getting one answer sooner."
    ],
    "topics": [
      "agents",
      "agent-harnesses",
      "codex",
      "developer-tools",
      "inference-cost",
      "inference-speed",
      "orchestration",
      "open-source"
    ],
    "tools": [
      "OpenAI",
      "Codex",
      "Codex app",
      "Codex Cloud",
      "Codex Monitor",
      "Codex for iOS",
      "Responses API",
      "AGENTS.md",
      "app server",
      "GPT 5.6 (Sol, Soul, Terra, Luna)",
      "GPT 5.5",
      "GPT 5.3 Spark",
      "o1-preview",
      "ChatGPT",
      "Cerebras",
      "Open Code",
      "Open Claw",
      "Pi",
      "Droids",
      "Xcode",
      "JetBrains",
      "VS Code",
      "Vim",
      "Emacs",
      "GitHub",
      "Slack",
      "Terminal-Bench",
      "VNC"
    ],
    "quotes": [
      {
        "text": "Software ate the world. And then AI ate software. But now, what we're here to say is that the AI engineers are eating the world.",
        "at": "01:21",
        "url": "https://www.youtube.com/watch?v=pMggiOb18tc&t=81s"
      },
      {
        "text": "We're not building one system for Open AI and a second system that's simplified for developers. At every layer, we actually use the thing that we give to you.",
        "at": "13:37",
        "url": "https://www.youtube.com/watch?v=pMggiOb18tc&t=817s"
      },
      {
        "text": "I thought I was orchestrating. Really, I was polling. I was the scheduler, the router, and the memory.",
        "at": "19:26",
        "url": "https://www.youtube.com/watch?v=pMggiOb18tc&t=1166s"
      },
      {
        "text": "Now, I'm primarily constrained by attention. And unlike tokens or compute, I can't simply add more of it. So, the most important skill is today is deciding where to spend it.",
        "at": "21:20",
        "url": "https://www.youtube.com/watch?v=pMggiOb18tc&t=1280s"
      },
      {
        "text": "Models are advancing faster than the harnesses and organizations around them. Designing those things is the next engineering problem.",
        "at": "24:34",
        "url": "https://www.youtube.com/watch?v=pMggiOb18tc&t=1474s"
      }
    ],
    "words": 5365
  },
  "summary_url": "/#pMggiOb18tc",
  "transcript": {
    "html": "/transcripts/pMggiOb18tc.html",
    "txt": "/transcripts/pMggiOb18tc.txt",
    "vtt": "/transcripts/pMggiOb18tc.vtt"
  }
}