We have been working hard on the past two months implementing agents for different tasks. We started with Claude code, as we really liked working with hooks. However being vendor locked and having usage limit problems, we ended up implementing our own "runtime" that keeps the conversation structure while having hooks. Right now it only works with OpenAI but its designed to be able to incorporate Claude / Gemini down the road
We ended up open sourcing that runtime if anyone is interested:
While I like the self reflection from this article, I don't think his methodology adds up (pun intended). First there are two main axis where LLMs can make you more productive: speed & code quality. I think everyone is obsessed about the first one, but its less relevant.
My personal hypothesis is that when using LLMs, you are only faster if you would be doing things like boilerplate code. For the rest, LLMs don't really make you faster but can make your code quality higher, which means better implementation and caching bugs earlier. I am a big fan of giving the diff of a commit to an LLM that has a file MCP so he can search for files in the repo and having it point any mistakes I have made.
I strongly relate to this. I built my desktop computer and I run arch as my daily driver. For my phone: I just want a device that works out of the box. I don't want to think about it, or need to configure anything from it.
We ended up open sourcing that runtime if anyone is interested:
https://github.com/Vanclief/agent-composer