I created an agent using CodeInterpreterTool to run code generated by LLM. However, in some iterations, the CodeInterpreterTool fails in running the code, while the generated code itself is totally correct. I checked the crewai source code and found the following part:
cmd_to_run = f'python3 -c "{code}"'
exec_result = container.exec_run(cmd_to_run)
It seems like that when the code with double quotation mark inside will fail here. Are there any solutions?
I’ve found an cheap solution: tell the agent the error in the “goal” or “description” (or a way to format correctly) and the agent will format it correctly.
So in the end, it’s not a good idea to do things that way, it doesn’t work every time… The best is to use the tool: ‘FileWriterTool(directory=“./output“, overwrite=True) and then, in your task, use the option ‘expected_output : > write code to main.py’.
And add another task to execute the code
Have you tried output_file: main.py instead of expected_output : > write code to main.py. Does it work?
Both work, I think, and it’s hard to know which is better. But keep in mind that this is an example. For now, I’m using structured sentences for the ‘expected_output’.