How do you decide that an agent's output is good enough to stop iterating?

You may know the anegdote of Henry Kissinger. He asked Winston Lord to prepare a document. Each time Winston handed him the document, he asked, “Is this the best you can do?”

He finally started to read it only when Winston said "yes".

With AI it is easy to work like that, but hard to decide when to stop and we can fall in an infinite loop. Curious to know what your approaches on that point?