Amazon Bedrock has evolved into a full AI platform on AWS, covering foundation models from 18+ providers, managed RAG via Knowledge Bases, agent runtimes (classic Agents and the newer AgentCore), guardrails, fine-tuning, and observability. The guide covers when to choose Bedrock over direct provider APIs or self-hosting, how to use the Converse API as the default inference surface, RAG chunking strategies and vector store trade-offs, the difference between classic Bedrock Agents and AgentCore (microVM-based, framework-agnostic, MCP-compatible), cost management patterns including batch inference and prompt caching, and a 10-point decision framework for new projects. Key recommendations: default to Converse API, use Knowledge Bases unless retrieval is a core product feature, adopt AgentCore for production agents, add Guardrails from day one, pin model versions, and build evaluation harnesses before shipping.
Table of contents
What Amazon Bedrock IsWhen Bedrock Wins vs Direct APIs vs Self-HostedFoundation Models and the Inference APIsKnowledge Bases for RAGBedrock Agents and AgentCoreGuardrails, Custom Models, and Provisioned ThroughputSecurity, Observability, and Reference ArchitecturesPicking the Right PatternKey TakeawaysSort: