If this ain't bad enough, their philosophy is cherry on the cake
I think the main idea is that it's an "agent" that runs command line commands and then considers the output. It definitely helps sometimes to show a LLM the errors its code generated.