From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer

AI Engineer18m 31s
0 comments · 0 votesOpen discussionClose discussion
Sign in to join the discussion

    Video summary

    Rafal Wilinski explains why long-running agents need reusable skills rather than repeatedly rediscovering procedures. He identifies developer-centric interfaces, authoring effort and fragmented client support as barriers, and proposes governed MCP distribution as a shared access layer for technical and nontechnical users.

    Rafal Wilinski connects trace-based skill creation with Voyager's reusable Minecraft procedures. His approach distills successful and failed runs into actionable instructions, with a focused example improving a task's reported success rate from 36% to 100%; the result is an example, not a general reliability guarantee.

    Rafal Wilinski argues that organizations should retain procedural knowledge as frontier models change. Sharing vetted lessons across agents could create a compounding knowledge loop, but stored traces and generated skills still need privacy checks, security boundaries and review before wider use.

    Original YouTube thumbnailWatch on YouTube

    Share this page

    Rafal Wilinski in a light blue shirt beside SHARED AGENT SKILLS on a black background. Framed in blue with WWW.ARTIFICIAL-INTELLIGENCE.VIDEO, 7 October 2026 and duration 18m 31s.

    Rafal Wilinski proposes using MCP to distribute governed skills and distilling agent traces into reusable organizational knowledge.