They can (unintentionally) obfuscate what they’re doing.

I’ve seen the agent make scripts with commands that aren’t immediately obvious. You could unknowingly say yes when it asks for confirmation, and only find out later when looking at the output.