The RAG vs agentic search split you draw is the one most teams do not name explicitly, RAG assumes a curated corpus and agentic search assumes the open web, and those two lead to completely different eval strategies. Citation grounding is where I keep seeing the gap in production, the model can pull real URLs and still misattribute the claim, so the eval has to check the claim-to-source alignment, not just that a source exists. Would love to see a follow-up on how you score citation faithfulness at scale.