Agreed. I get the best outcomes out of agents when I use them as essentially autocomplete.

"Render what I was thinking"

I just double check to make sure it aligns