It actually is somewhat a limit of the technology. LLMs can't go back and modify their own output, later tokens are always dependent on earlier tokens and they can't do anything out of order. "Thinking" helps somewhat by allowing some iteration before they give the user actual output, but that requires them to write it the long way and THEN refactor it without being asked, which is both very expensive and something they have to recognize the user wants.
Coding agents can edit their own output - because their output is tool calls to read and write files, and so it can write a file, run some check on it, modify the file to try to make it pass, run the check again, etc
