Show HN: AutoLearn Skills for self-improving agents

(autolearn.dev)

27 points | by toobulkeh 3 days ago ago

11 comments

  • toobulkeh 2 days ago

    I vibe coded this at a hackathon a few weeks ago. The solution isn’t stable yet, but it’s similar to Claude Skills with the benefit of only needing an MCP client.

    I open sourced it too.

    It has a bit to go—like automatically monitoring reasoning traces and security—but I thought this would be easier to use than trying to maintain your own skills.md collection if the server handled more of it for you.

    I’d be curious if anyone thinks I should keep working on it. I’d love someone to collaborate with on it!

  • stpedgwdgfhgdd 2 days ago

    I sometimes have to decide between running Claude Code headless with a prompt or let CC generate an application based on the prompt. (coining: Codification Threshold tradeoff)

    The tradeoff is that the prompt is non-deterministic, but also more flexible. It can handle unknown situations. The generated code executes much faster and reliable for known situations.

    If auto-learn would reduce the non-deterministic nature, i.e. chance of failing, that would speed up software development.

  • snowfield 2 days ago

    This is a very interesting concept. I have a few questions though.

    Even if the agent has mcp tools avaible, you can't guarantee that it would necessarily use them right? I imagine that you'd have a lot of new tools that do nearly the same things.

    Also how does this scale? If the agent has like 300 tools.

    • toobulkeh 2 days ago

      I don’t know (yet) but some ideas might be: 1. A skill that searches skills 2. MCP that returns only relevant skills based on context 3. Separate concerns through sub agents that each have their own autolearn server

  • drdeca 2 days ago

    How does it determine whether the code it has for the skill, failed? (In order to determine that it needs to improve the code for that skill, I mean)

    • toobulkeh 2 days ago

      Definitely needs some work here. That isn’t done (yet). Great feature request! I think it would be some kind of testing logic or feedback loop. Or both!

  • personjerry 2 days ago

    I don't understand, it'd be nice to see even a single example of the application