THIS so much. People are like "why human supervision when we can have agent supervsion" and always respond > look if you don't trust the LLM to make the thing

THIS so much. People are like "why human supervision when we can have agent supervsion" and always respond > look if you don't trust the LLM to make the thing right in the first place, how are you gonna PROBABLY THE SAME LLM to fix it? yes I know multiple passes improves performance, but it doesn't guarantee anything. for a lot of tool you might wanna call, 90% or even 99% accuracy isn't enough