I'm still experimenting with LLMs. I made a self-improving agent from scratch and wrote about what I learned: https://chris.eidhof.nl/post/self-forking-agents/