Treat the LLM as a brain, not a source of facts. A practical guide to multi-agent LLM systems: prompts, orchestration, RAG, vector search, evals, and inference.