{
  "video": {
    "id": "p7gaxmCUiIU",
    "title": "Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase",
    "duration": 3113,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "p7gaxmCUiIU",
    "title": "Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase",
    "one_liner": "Browserbase's \"auto browse\" loop is a version of Karpathy's auto research. A doer agent runs a browser task while a teacher agent reads its traces and builds up a strategy.md across iterations. The loop ends in a portable skill.md, which made browser agents up to 16x cheaper and much more reliable without touching model weights.",
    "summary": "Srivastava argues that agent reliability is an engineering problem in the harness, tools and skills, not in the model. Today's browser agents are \"geniuses that never learn\" because every run starts from zero. He shows auto browse: two Claude Agent SDK agents (doer and teacher) iterate on a live task, write what they learn to strategy.md and stop when the result converges into a reusable skill.md. Live demos on Google Flights and OpenTable show costs falling from $12 to $3 and from $6.22 to $0.51. The agents found tricks like encoding the whole query in URL parameters and avoiding the OpenTable endpoints that trigger anti-bot blocks. He says the method carries over to any domain that has a replayable environment and a verifiable outcome.",
    "key_points": [
      "Auto browse runs two Claude Agent SDK agents. The doer drives the browser through the browse CLI (open, snapshot of the accessibility tree, screenshots, clicks). The teacher reads the doer's traces, screenshots and logs and appends what it learned to strategy.md, which carries over to the next iteration. Converged output is a skill.md for future agents.",
      "The loop is given no iteration count, success-rate target or token/time budget. Following the 'bitter lesson', the agents decide when they have converged, and the teacher 'calls it a day'. Pass/fail is self-evaluated and usually binary, with few guardrails.",
      "Google Flights demo (Opus 48 as both doer and teacher): iteration 1 learned to skip the calendar widget and use a query parameter. Cost fell from $12 to $3 (4x) by iteration 2. Iteration 3 overcorrected and failed. The final skill found an 8am business-class SFO–NYC flight for $649 with a carry-on.",
      "OpenTable demo: the 'access denied' block turned out to be environmental. The agent learned to build a search URL with latitude/longitude/metro ID and never to call /r or reload /s. Cost fell from $6.22 to $0.51 and time from 189s to 27s.",
      "Customer results range from 3x to 16x cheaper. Some customers have built pipelines that generate skills on the fly, open PRs automatically, and start a healing loop when a skill fails in production.",
      "browse.sh is described as 'yellow pages for browser agents': about 500 site skills generated by the same loop, served free through the browse CLI (Browserbase pays the token costs), stored in Vercel Blob. Users can also request a new skill to be generated.",
      "When optimizing for latency, the loop sometimes drops the browser entirely, reverse-engineers the site's API and gets a 500 ms solution. Sonnet run for more iterations can match Opus run for fewer, at lower total cost. One loop costs about $30–40, $100 at most, and that cost is amortized.",
      "Prompt-injection defence has three layers: guardrails before the query reaches the harness, a security-reviewer agent that checks skills before they are published to browse.sh, and infrastructure controls (Browserbase safe mode restricting URLs and actions, network firewall, allow-once/always-allow approvals)."
    ],
    "takeaways": [
      "Have agents learn from their own production traces with a doer/teacher loop that writes to a persistent strategy file, instead of fine-tuning per task. Ship the converged result as a portable skill.md that works in any agent or SDK (Claude Code, Stagehand, OpenClaw).",
      "Before applying auto-research outside browsing (coding conventions, research sources, internal ops tools), make sure you have (a) a replayable, 'grindable' environment and (b) a verifiable or quantifiable outcome. Investing in the environment pays off.",
      "Choose the loop's optimization target on purpose (success only, latency, number of turns, or cost), and use evals to decide between one skill per task and one skill for all tasks.",
      "Don't re-run auto research daily. Run a cron job (e.g. a daily 9am check on Anthropic Managed Agents) or react to production failures, then trigger a healing loop that opens a PR for a human to review.",
      "Try running a cheaper model like Sonnet for more iterations instead of Opus for fewer; it can reach the same result at lower total cost."
    ],
    "topics": [
      "agents",
      "browser-agents",
      "skills",
      "auto-research",
      "self-improvement",
      "reliability",
      "evals",
      "prompt-injection"
    ],
    "tools": [
      "Browserbase",
      "browse CLI",
      "browse.sh",
      "auto browse",
      "skills.sh",
      "Claude Agent SDK",
      "Claude Code",
      "Codex",
      "Opus",
      "Sonnet",
      "Fable",
      "Stagehand",
      "OpenClaw",
      "Hermes",
      "Anthropic Managed Agents",
      "Vercel Blob",
      "WebMCP",
      "Google Chrome",
      "DSPy",
      "GEPA",
      "Microsoft",
      "Google Flights",
      "OpenTable",
      "Amazon"
    ],
    "quotes": [
      {
        "text": "these browser agents, I call them geniuses that never learn",
        "at": "05:50",
        "url": "https://www.youtube.com/watch?v=p7gaxmCUiIU&t=350s"
      },
      {
        "text": "this was like my moment of like this is the mythos moment for browser agents you know like just reverse engineering the hell out of things",
        "at": "17:32",
        "url": "https://www.youtube.com/watch?v=p7gaxmCUiIU&t=1052s"
      },
      {
        "text": "This is yellow pages for browser agents right it has details about how to do any task on any website on the web.",
        "at": "18:56",
        "url": "https://www.youtube.com/watch?v=p7gaxmCUiIU&t=1136s"
      },
      {
        "text": "sometimes these auto research loops are so smart they realize that browsers are not needed at all and they just reverse engineer the API",
        "at": "40:08",
        "url": "https://www.youtube.com/watch?v=p7gaxmCUiIU&t=2408s"
      }
    ],
    "words": 9045
  },
  "summary_url": "/#p7gaxmCUiIU",
  "transcript": {
    "html": "/transcripts/p7gaxmCUiIU.html",
    "txt": "/transcripts/p7gaxmCUiIU.txt",
    "vtt": "/transcripts/p7gaxmCUiIU.vtt"
  }
}