{ "video": { "id": "buHC7bQE1X4", "title": "Get Out of the Model's Way — Kevin Hou, Google Antigravity", "duration": 1141, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "buHC7bQE1X4", "title": "Agents Built an OS Kernel That Runs Doom — Kevin Hou, Google DeepMind", "one_liner": "Kevin Hou's rule for building around LLMs is to \"scale with intelligence\": pick product primitives that get better as the model does. Antigravity 2.0 bets on three of them for 2026 (dynamic sub-agents, sidecar triggers and generative UI), and he backs this with a Gemini 3.5 Flash agent team that built an OS kernel running Doom using 93 sub-agents in 12 hours for under $1,000.", "summary": "Kevin Hou leads part of the engineering team on Google's Antigravity. He argues that LLMs can now be the 'star player', so products should 'give Messi the ball and get out of the way'. He walks through the primitives each era added: autocomplete and chat in 2022, agents with MCPs and permissions in 2024, the agent manager with skills, hooks and artifacts in 2025. He says 2026 is the era of agent teams, built on dynamic sub-agents, sidecars and generative UI. He shows this with Antigravity 2.0's /teamwork mode, the OS-kernel-plus-Doom run, and an internal DeepMind workflow for side-by-side evals.", "key_points": [ "Antigravity launched in November 2025. It then extracted its agent into the Antigravity CLI, and Antigravity 2.0 (launched at Google IO) split the IDE and the agent manager into two separate apps. His framing: 'the IDE is to the agent manager what the debugger was to the IDE.'", "Past 'battle scars' from scaling with intelligence: giving AI a terminal (made safe by permission systems and smarter models), and removing the chat sidebar from Windsurf in favour of an agent only, which drew angry user backlash at the time.", "Gemini 3.5 Flash (launched in April) learned to manage agent teams through the research-product flywheel with Antigravity. It is 'not just good at executing tasks, it's actually really good at leading teams', and it is faster and cheaper.", "Agent teams are in public preview and started with the /teamwork slash command. A lead agent asks clarifying questions, then dynamically generates specialised sub-agents of arbitrary number (frontend, backend, infra, QA, design). The lead agent can choose a different model for each sub-agent.", "The hero run built a complete OS kernel from scratch that plays Doom. It took 93 sub-agents, 12 hours, 15,000 requests and 2 billion tokens, and cost under $1,000. Other builds were a raw photo editor that runs in the browser and a messaging app, each using hundreds of sub-agents over almost half a day.", "Internal side-by-side eval workflow: researchers automated 90% of it with natural language plus a skills file. The agent computes the delta between control and experiment, spins up a research specialist that proposes 100 hypotheses, runs one sub-agent per hypothesis in parallel, then produces a report and an interactive generative UI. It replaced Jupyter notebook 'elbow grease' and hand-built async agent pools, judges and data pipelines, and now takes minutes.", "Sidecars are a new plugin protocol: long-lived processes that let the model listen to the outside world and set its own triggers (SMS, webhooks, cron jobs, GitHub PRs). Antigravity's scheduled tasks already run on sidecars, and the spec will be released later this summer.", "Generative UI: Gemini Flash on Antigravity runs at almost 900 tokens a second, which he says is 10x faster than many other frontier-model experiences. That speed lets on-demand UIs (kanban boards, timelines, charts, tables, even Doom) render inline in the conversation instead of coming from templates or HTML files. His claim: 'human written specialized UIs are kind of dead.'" ], "takeaways": [ "For each primitive in your agent product, ask how it will scale with the model's intelligence. Shipping features is now easy; the hard part is choosing which ones will still pay off when the next model arrives faster, better and cheaper.", "Be willing to remove familiar features that users love, such as a chat sidebar or an IDE-centric workflow, when the model can support a better agent-first path, even if users push back.", "Let the orchestrating agent configure, prompt and pick models for its sub-agents dynamically, rather than hard-coding fixed roles.", "Automate research and eval analysis by combining a skills file, fan-out sub-agents per hypothesis, and generated interactive UI for the results, instead of hand-building agent pools, judges and pipelines.", "Prefer generic, model-driven primitives (sidecar triggers, generative UI) over heavy infrastructure and fixed 'mechanical UIs' that are 'fixed in plastic'." ], "topics": [ "agents", "multi-agent orchestration", "sub-agents", "generative ui", "developer tools", "evals", "product strategy" ], "tools": [ "Antigravity", "Antigravity CLI", "Antigravity 2.0", "Gemini 3.5 Flash", "Google DeepMind", "Google IO", "Windsurf", "MCP", "Jupyter", "Chrome DevTools", "GitHub", "Doom" ], "quotes": [ { "text": "It's called Give Messi the ball and get the heck out of the way.", "at": "00:53", "url": "https://www.youtube.com/watch?v=buHC7bQE1X4&t=53s" }, { "text": "We believe that the IDE is to the agent manager what the debugger was to the IDE.", "at": "05:59", "url": "https://www.youtube.com/watch?v=buHC7bQE1X4&t=359s" }, { "text": "So we hypothesize that human written specialized UIs are kind of dead.", "at": "16:04", "url": "https://www.youtube.com/watch?v=buHC7bQE1X4&t=964s" }, { "text": "And that creates a product experience that is not fixed in plastic.", "at": "17:22", "url": "https://www.youtube.com/watch?v=buHC7bQE1X4&t=1042s" } ], "words": 3950 }, "summary_url": "/#buHC7bQE1X4", "transcript": { "html": "/transcripts/buHC7bQE1X4.html", "txt": "/transcripts/buHC7bQE1X4.txt", "vtt": "/transcripts/buHC7bQE1X4.vtt" } }