Except RAG which is arguably the most important use case, what do you need ? Unreliable agents that are not capable of rigorous reasoning ??