A practical architecture review of how to design a production-grade AI agent platform on AWS—starting simple, then adding RAG, vector search, agents, memory, MCP tools, authorization, guardrails, observability, and evaluation only when the requirements justify them.