Cline w/claude, Cursor, or copilot agents. There are probably better answers for how this works, but with copilot agents it breaks work down into small pieces and feeds it back to the LLM so it itterates on smaller components isntead of trying to do the job all at once. In lay terms: I think the models understand the constraints of the output windows size and try to get entire thing done in one sweep, they currently have no concept of history or memory, that's where tools come in to help break the work down and keep feeding it pieces.