Agentic RAG: The Complete Production Guide Nobody Else Wrote

Three months into a contract with a mid-sized insurance company, I was sitting across from their CTO watching their "AI knowledge base" answer questions about their own products. The system retrieved the right documents 90% of the time. But on anything involving multi-part questions, comparisons, or anything that required checking two sources together, it fell apart. Their agentic RAG system wasn't agentic at all. It was a fixed pipeline wearing an agent costume, and it was costing them about $4,200 a month in API calls to produce answers that were wrong 62% of the time on complex queries. That project is what pushed me to formalize what I now call an agentic RAG system the right way. I've since deployed some form of this architecture across 38 of my 109 production AI systems, and the patterns I'm about to share are hard-won. This guide covers what most agentic RAG articles skip: real chunking decisions, embedding model comparisons, the four failure modes that will definitely hit you i