[Bug]: malloc issue when using tools and agent run in a loop

[marklysze]

Describe the bug

I was trying to run an agent with tools using the agent's run method while in a for loop and it was causing some memory allocation issues during the loop. It wasn't consistently at any spot but did happen on around iteration 8 and 12.

Using the GoogleDriveToolkit and one of the tools in it: download_file_from_drive

Here's an output of it - which has some printouts of the run events (the 11/98 was iteration 11 of 98):

[11/98] type='using_auto_reply' content=UsingAutoReplyEvent(uuid=UUID('d286aab8-9e86-4f36-ba53-4a80720a6314'), human_input_mode='TERMINATE', sender='user', recipient='download_agent')

[11/98] type='tool_call' content=ToolCallEvent(uuid=UUID('4f014cd7-ee70-4606-a074-b6384958cf22'), content=None, sender='download_agent', recipient='user', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[ToolCall(id='call_KtBxbrw92HrWnPAH1d0PcXEn', function=FunctionCall(name='download_file_from_drive', arguments='{"file_info":{"name":"Support Letter.pdf","id":"1UR69JhOMQsNqDOYX1qSndVQvwHT11cFo","mimeType":"application/pdf"},"subfolder_path":"test_folder"}'), type='function')])

[11/98] type='execute_function' content=ExecuteFunctionEvent(uuid=UUID('d00a1b61-6154-4486-81c2-aa757b9f2197'), func_name='download_file_from_drive', call_id='call_KtBxbrw92HrWnPAH1d0PcXEn', arguments={'file_info': {'name': 'Support Letter.pdf', 'id': '1UR69JhOMQsNqDOYX1qSndVQvwHT11cFo', 'mimeType': 'application/pdf'}, 'subfolder_path': 'test_folder'}, recipient='user')

Python(71119,0x324013000) malloc: Incorrect checksum for freed object 0x16a80c400: probably modified after being freed.
Corrupt value: 0xaf3664d4a0779dca
Python(71119,0x324013000) malloc: *** set a breakpoint in malloc_error_break to debug

@davorrunje: If you are refactoring the tools and having a more centralised way of adding tools to agents, you can consider taking the functions parameter off ConversableAgent as that's another interface for adding functions and I think the combination of all these interfaces for adding tools is producing some level of inconsistency.

Steps to reproduce

No response

Model Used

No response

Expected Behavior

No response

Screenshots and logs

No response

Additional Information

No response