Running a command in a shell is a string of text. LLMs produce text and it's easy to write a program that then executes that text as a process. I don't see what's magical about it at all.