◇ Could this help me?
Yes, and this is a useful counterweight to the loop content that circulates without caveats. The critique is correct on both grounds: a specification detailed enough to eliminate real questions is usually more work than the task itself, and when an agent hits an unpredicted case in a loop it guesses rather than asking, so the failure arrives buried in finished-looking work. The cost point is also fair, since iterating against assumptions is exactly where token spend goes without producing value. General practice, and the honest nuance is that loops are not useless, they are appropriate for narrow, well-specified, verifiable tasks where a wrong guess is cheap to detect.
I am deciding whether to run a task as an autonomous loop or keep myself in the loop. Take the task I describe below and assess it honestly against the things that make a loop work: is the specification complete enough that no real questions will arise, is there an automatic verifier that can tell success from a plausible-looking failure, and is a wrong guess cheap to detect and reverse. Tell me which parts of the task fail those tests and would need a human decision, and propose a split where the well-specified parts run unattended and the ambiguous parts come back to me.