Humans understand that missing information is also information. LLMs don't because they were literally built to "fill the blanks". But us humans know sometimes blanks must be left alone and cannot be filled.
Sub agents were invented to solve for this: before generating more tokens, first spin up another LLM to figure out what to do with the missing info. This is a fallacy though because humans would know that we should avoid the work completely until we have further information. LLMs entire purpose is to generate the next token, so they will put in the work. This violates nature's "path of least action" principle.
We end up with potentially millions of tokens generated that lead to nowhere (best case) or they lead to a misleading hallucination (worst case) while it also costs electricity and time. Meanwhile, a human could walk in the park to feed the ducks at the same time, achieving MORE.