I like the idea!

[luisferrassini]

So basically I've been trying to create the thing you already created. I don't have many llm tokens to spend at this so it's quite unfinished but maybe you could take a look and see how they would integrate? The thing I struggle the most is getting the harnesses setup correct (because I run out of tokens so fast and my agents get all confused leading to junk artifacts all over the place) so I can just do the development flow I like the most, which is kinda like: planning -> reviewing -> enforced tdd -> review -> commit push done
I like the superpowers flow but feels very heavy on the tokens sometimes.
Here's the repo: https://github.com/luisferrassini/agent-helm/tree/agent-helm-milestone-1

[luisferrassini]

Also I'll just be honest and say that I vibecoded the whole thing because I've never worked with the go language before, but at least see if my idea could make sense together with your project

[codeaholicguy]

Glad to hear that you also have the same idea and vision, will definitely take a look at your solution.

[luisferrassini]

I wrote the issue late night 2 days ago, let me say I was a bit tired, so I'll rewrite my thoughts.
So, basically my vision was: Go to a project, install agent-helm (my project) and then it gives you a command to run the harness of your choosing with a given profile, that could be "vanilla" (no skills/plugins/anything, just the raw installation of the harness), or with some workflow predefined like superpowers only, or mattpocock skills, or even ai-devkit profile.
I find it very hard to test these different kinds of setups on real projects that I work on. These files get all scaterred around and makes it worse when you have to change harnesses constantly because they keep running out. Would be nice to have a unified app that deals with this. I also discovered now about oh-my-claude, oh-my-openagent, oh-my-opencode-slim, and so on, and these could make it into my harness setup idea as profiles as well. I haven't tried your ai-devkit but the things I've seen on your blog posts honestly look fire.
TLDR just want to think of a solution to my problem of setting up ai coding agents for projects.

[codeaholicguy]

Hi @codeaholicguy,

Thanks for sharing more context. This makes a lot of sense.

One of the problems I am trying to solve with ai-devkit is exactly this: having one setup/config layer that can work across different agents, instead of scattering files across projects and then having to redo the setup whenever you switch harnesses.

You can take a look at this part of the docs: https://ai-devkit.com/docs/9-agent-setup/

For your use case, I think agent-helm could potentially integrate with ai-devkit and leverage the agent template concept. The template already supports configuration for skills and MCP servers, and it works across different agent types. For example:

environments:
  - claude
  - codex
skills:
  - registry: codeaholicguy/ai-devkit
    skill: structured-debug
  - registry: codeaholicguy/ai-devkit
    skill: dev-lifecycle
  - registry: obra/superpowers
    skill: brainstorm
  - registry: mattpocock/skills
    skill: grill-me
mcpServers:
  memory:
    transport: stdio
    command: npx
    args:
      - -y
      - "@ai-devkit/memory"

Then you can run 2 commands to set the agent up.

ai-devkit init -t agent.yaml
ai-devkit install

I can look into adding more types of cross-agent config when there are requests that make sense.

Overall, I think we are trying to solve a very similar pain point: making AI coding agent setup easier to reproduce, switch, and test across real projects.

Let me know if you have any further questions.

[luisferrassini]

Thank you very much for your response.

So I don't really want to use the agent-helm project, I just wanted to solve the harness/agent setup problem. If you already solved this issue with your agent-setup, that's great news for me!

I'll try it out when I have the chance and let you know. Cheers!