Exactly why you can't remove humans in the loop to assess that the solution is not only correct (which LLMs are quite bad at, once concurrency, logic, etc are involved), but also elegant, maintainable, etc