I’ve had great success with this method already and even built a serialization CLI tool [1]. However, with O3-Mini, I’ve noticed something new. O1-Mini is both capable and fast. My previous workflow involved waiting quite a bit for the LLM to respond, so I didn’t mind spending time on: 1. Collecting the relevant code 2. Fixing minor issues in the generated code
But now, most of my time is spent making small fixes to the AI-generated code. Many of these “bad responses from the AI” could be avoided if the AI had insight into the runtime values of the variables in the code.
I believe it would be very beneficial to have IDE-like, debugger-mode interactions with the LLM. Imagine sending the LLM all of the variable values and letting it pause, continue, step over, and step into the code.
In short, I’m asking if anyone is working on an agent that uses the IDE like a human would.
[1] http://github.com/bodo-run/yek