Ask HN: How to separate agent harness from code repo?

I’m interested in any approaches or ideas for how to keep agent harness and code separate, like ideally separate repos.

Spec driven development is going really well for me, along with a swarm of agents and skills etc.

However, I’m seeing .Claude/memory/, docs/, .specify/*, hooks and extensions, context management tools etc becoming as much an overhead to manage as the codebase itself sometimes.

Also, it’s all mixed in together, in the same repo. Surely this is an anti-pattern.

Are there better ways to do this (yet)? Or is this the next AI frontier to evolve?

2 points | by FrankRay78 1 hour ago

2 comments

  • verdverm 1 hour ago
    I've been using and taking inspiration from Scion[1], where repos + harnesses + agents = working env

    There are global files, and repo files, but we are missing a good solution to the middle, which (imo) is best served by a proper versioning and dependency management system. (like CUE mods, this can probably fill the gap by generating config and embedding the markdown as part of that, one liner to write out an imported module into the current repo)

    I don't think we really need anything new here (ai hype wants to reinvent everything), we have plenty of tools and prior art to leverage.