You shouldn't be downvoted - LLMs could in theory be deterministic, but they currently are not, due to how models are implemented.
It is absolutely workable, current inference engines are just lazy and dumb.
(I use a Zobrist hash to track and prune loops.)