We're trying too hard to have one model do it all. If we coordinate multiple models + other tools (ala ReAct pattern) we could make the systems more resistant to prompt injection (and possibly other) attacks and leverage their respective strengths and weaknesses.
I'm a bit wary of tool invocation via python code instead of prompting the "reasoning" LLM to teach it about the special commands it can invoke. Python's a good crutch because LLMs know it reasonably well (I use a similar trick in my project, but I parse the resulting AST instead of running the untrusted code) so it's simpler to prompt them.
In a few iterations I expect to see LLMs fine tuned to know about the standard toolset at their disposal (eg. huggingface default tools) and further refinement of the two-tiered pattern.