I would simply replace LLM by agent in your reasoning, in the sense that you'll need a strong preprocessing step and multiple iterations to exploit such complete specs.

There is sense in your words. Especially in the context of the modern day vocabulary.

I though about the concept of this ort of methodology before "agent" (which I would define as "sideeffects with LLM integration") was marketed into community vocabulary. And I'm still rigidly sticking to what I consider "basics". Hope that does not impede understanding.