{
  "video": {
    "id": "u-o0sW9nwmk",
    "title": "From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer",
    "duration": 1111,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "u-o0sW9nwmk",
    "title": "From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer",
    "one_liner": "Use frontier models while you still have them to turn agent runs, especially failed ones, into vetted skills, then serve those skills company-wide over a remote MCP server so every agent and employee starts from the best known path; at Runlayer this took a Sonnet 4.6 task from a 36% to a 100% success rate.",
    "summary": "Wilinski, a founding engineer at Runlayer who previously ran Zapier's agents work, argues that skills (instructions an agent loads through progressive disclosure) are the most accessible way to improve agents. Longer-running agents can go far down a wrong path, so the starting direction matters. He names three problems with skills: they are local and built for developers, nobody takes the time to write them, and many clients don't support them. His answer is to distribute skills over MCP and generate them automatically from agent traces, an idea he takes from Voyager's skill library. Pooling this across every agent in a company gives what Runlayer calls an 'organizational flywheel of self-improvement', which he presents as a moat in a world where models are a commodity.",
    "key_points": [
      "Definition: a skill is an instruction the agent reads only when it judges it relevant. This is progressive disclosure: the client first sees only a short description, and the agent pulls the full text into context when it needs it, which changes what it does next.",
      "Citing the METR benchmark (transcribed as 'MEETER'), agents can work on their own for longer and longer. The catch is that a long-reasoning agent can make hundreds of tool calls defending a false thesis, so the first trajectory matters, and skills steer it. Without them you waste time, tokens and money.",
      "Three problems with skills: (1) they are local and aimed at developers (markdown, JSON, CLI commands, Git repos), which leaves out non-technical staff; (2) usually nobody has written the plan down; (3) not every LLM client supports skills, and departments like marketing, HR, legal and finance often use ChatGPT. He also jokes that `npx skills` installs 'free prompt injections, a wild script with root privileges, and someone else's API key', and says Runlayer scans thousands of skills a day.",
      "He calls 'MCP is dead' a misreading: skills convey how to do something, while tools give capabilities. MCP can also act as the distribution layer for skills. A remote MCP server becomes the company's source of truth, enforces policy, and serves the best version of a skill by request and role. Runlayer has run this internally for about 6 months, and a 'Skills over MCP' working group is heading the same way. The spec leans towards MCP resources, but client support for resources is patchy, so Runlayer exposes skills as tools.",
      "Voyager (2023, GPT-4, Minecraft) improved without changing model weights or prompts, using a skill library: each new discovery was stored as reusable code made of tool calls, such as the recipe for a workbench or diamond pickaxe. The lesson he draws is that research should become reusable procedural knowledge.",
      "Runlayer's distiller takes the trace of a successful or interesting run and asks a frontier LLM to extract a skill. Failed runs are the richest source because they expose missing safeguards ('fuses' in the transcript), libraries and edge cases. Successes make a skill more reliable and failures make it more robust, all autonomously and asynchronously.",
      "Result: running a Chromium fork inside AWS Lambda with Sonnet 4.6 succeeded 36% of the time. With self-improvement on, one run found a solution, the distiller turned it into a reusable algorithm, and with that skill in every later run the success rate went to 100%. Good skills also narrow the search space, so the agent makes fewer tool calls, which saves tokens and cost.",
      "Organizational flywheel: Runlayer groups similar runs, extracts a skill, checks it for prompt injection, PII and more, adds it to a skills library, and serves it through its MCP gateway to every agent and employee in the company."
    ],
    "takeaways": [
      "Write skills for the dull, company-specific routines no one wants to rediscover, such as the script sequence for a roadmap, the wording for a client who is about to leave, or the commands to restart a database or roll back production. Missing context like this is, in his view, why many AI pilots fail.",
      "Serve skills from a central remote MCP server, exposed as tools rather than resources for now, so non-technical staff in any client can use them and one owner controls policy, versions and role-based access.",
      "Log agent traces and distill skills from them automatically, mining failed runs as well as successes, and refine existing skills as new runs come in.",
      "Never install third-party skills blind. Scan them for prompt injection, PII, dangerous scripts and embedded keys before they enter the shared library.",
      "Use the best model available now to discover and record procedures as skills. This protects you against models being withdrawn or getting worse, and lets cheaper models perform better."
    ],
    "topics": [
      "agents",
      "agent skills",
      "self-improving agents",
      "mcp",
      "knowledge management",
      "progressive disclosure",
      "enterprise ai",
      "agent security"
    ],
    "tools": [
      "Runlayer",
      "Zapier",
      "METR benchmark",
      "MCP (Model Context Protocol)",
      "Skills over MCP working group",
      "npx skills",
      "Voyager",
      "Minecraft",
      "GPT-4",
      "Sonnet 4.6",
      "Opus",
      "Fable",
      "ChatGPT",
      "AWS Lambda",
      "Chromium"
    ],
    "quotes": [
      {
        "text": "Valuable skills are those rituals that no one wants to rediscover.",
        "at": "04:31",
        "url": "https://www.youtube.com/watch?v=u-o0sW9nwmk&t=271s"
      },
      {
        "text": "Skills can convey knowledge about how to do something, while tools give us capabilities.",
        "at": "07:48",
        "url": "https://www.youtube.com/watch?v=u-o0sW9nwmk&t=468s"
      },
      {
        "text": "research should be transformed into reusable procedural knowledge.",
        "at": "11:28",
        "url": "https://www.youtube.com/watch?v=u-o0sW9nwmk&t=688s"
      },
      {
        "text": "advanced intelligence is a rented or borrowed resource.",
        "at": "14:07",
        "url": "https://www.youtube.com/watch?v=u-o0sW9nwmk&t=847s"
      }
    ],
    "words": 3491
  },
  "summary_url": "/#u-o0sW9nwmk",
  "transcript": {
    "html": "/transcripts/u-o0sW9nwmk.html",
    "txt": "/transcripts/u-o0sW9nwmk.txt",
    "vtt": "/transcripts/u-o0sW9nwmk.vtt"
  }
}