Need help avoiding agent calling tools with json

I've noticed that - particularly weaker models - often call tools using json (in addition to writing code).

e.g.:
```
Thought: Let me try to get the sections using more specific search terms and the bm25_retriever to see the actual content first.

Code:
```py
print("=== SEARCHING FOR FIRST THREE SECTIONS CONTENT ===")
sections = bm25_retriever(query="Need for the Proposed Development Economic Benefits Recreational Benefits Wind Farm introduction",
num_snippets=5)
print(sections)
```[{'id': 'call_5', 'type': 'function', 'function': {'name': 'python_interpreter', 'arguments': 'print("=== SEARCHING FOR FIRST THREE SECTIONS
CONTENT ===")\nsections = bm25_retriever(query="Need for the Proposed Development Economic Benefits Recreational Benefits Wind Farm
introduction", num_snippets=5)\nprint(sections)'}}]
```
Are there any tips on avoiding this?

The report provides no repository files or tests; begin with the shown python_interpreter and bm25_retriever tool-call example. Reproduce the JSON-formatted call and trace how tool calls are generated and parsed. Done should be a defined, tested way to prevent or handle this malformed format.

Written by the indexing model from the issue text.