Comment by fb03

I've created a toy coding agent called "caducode". More of a thought experiment that materialized into a little something.

The philosophy behind it is: instead of providing a bunch of tools to the LLM, you simply provide a single tool: run_python(). The Agent just generates code to do whatever it needs, to inspect local files, to carry edits, to run commands.

https://github.com/flipbit03/caducode

It worked surprisingly well, even with a very small 30b local model.