Found this to be true across different models: The agent starts a python REPL to inspect some object. For any subsequent commands they automatically enter the REPL and try to execute terminal commands ...