Very poor performance writing code, possibly due to structure?

[mashdragon]

I followed the Function Calling Agent guide and created a function that lets the agent create save text to a file. Unfortunately, when I ask it to write some code and save it to a file, the results are really bad even on large models. I have a feeling that it's because of the format that llama-cpp-agent internally uses (which I don't know what that is). If it's JSON for example, that is understandable the quality is bad, because it wasn't trained on writing code using JSON only. It was trained on writing code with four spaces for each indent etc.

So I wonder if there are other ways for the agent to specify arguments for functions or a "think out loud" kind of approach so the input space matches the input space the LLM was trained on.