And who knows, perhaps a final filter to select "unnatural shapes that look like pyramids, platforms and plazas". :-) It really doesn't seem hard to believe that any of this is outside the capabilities of current gen LLMs. Including generating a plan to run multiple agents concurrently.