Fantastic question. Our goal is to use LLMs to help create the loops, but to rely on code for the majority of (predictable) automation. LLMs make for useful blocks, and oftentimes you can fallback to GPT 3.5 for most structured tasks. With that said, this launch definitely hit some GPT-4 rate limits :)

Also, we hope that an open-source version will allow folks to do more with their own keys.