Lesson 1: Agentic RAG is not always the right default

The single most important lesson. Agentic RAG should not be seen as a universal replacement for traditional RAG. It does offer better adaptability and multi-step reasoning, but it also adds coordination complexity, latency, and computational cost. For simple fact retrieval or well-scoped queries, a modular RAG pipeline gives enough performance at much lower overhead. The guidance is to adopt agentic designs selectively, guided by task complexity. This is the same instinct as "find the minimal effective system" that runs through the context-engineering and multi-agent topics.