...this is assuming humans stay in-the-loop.

Otherwise, for efficiency, LLMs could adopt something akin to:

Lambda Calculus in 383 Bytes

http://justine.lol/lambda/