jtor

I've reached a point where tweaking system prompts, tool docstrings, and Pydantic data type definitions no longer improves LLM performance (tool args accuracy and speed). I'm considering a multi-agent setup with smaller…