Overcoming Hard-Question Failures in RAG Using Agentic Design Patterns

I reviewed post-mortems from teams deploying Retrieval-Augmented Generation (RAG) in production and found a consistent failure mode: RAG handles simple questions well but breaks down on ambiguous quer