Multi-agent LLM systems fail for one of three root reasons, not a random assortment of bugs. The MAST taxonomy, what each looks like in a real trace, and where it shows up in a customer's numbers.
Every RAG evaluation guide covers the same five metrics and stops at the CI log. Here's the concrete cost math, real thresholds, and the customer-facing gap none of them touch.
Most golden dataset guides assume one team scoring one flat input-output pair. An AI agent needs the full trajectory, tagged per tenant, and tied to the number your own customers see.
G-Eval scores how good a reply sounds using an LLM judge and a chain-of-thought rubric. For an agent that plans and calls tools, that's only half the job.
Shadow testing runs a candidate AI agent version against real traffic before it ships, but comparing two non-deterministic outputs and stopping duplicate tool calls take real engineering the generic guides skip.
RAG vs agentic RAG comes down to one added loop, and that loop is where production failures start. What breaks, how to test for it, and the real cost math.