Agents in production: tools, state, protocols
Lesson 3 of 4 in Deploying FMs, RAG, and Agents.
An agent in production is four things wired together: a foundation model, instructions (the task definition), tools it may call, and state that survives across turns. On AWS that wiring comes in two strengths. Bedrock Agents is the managed, declarative version: you write instructions, attach action groups (tool schemas backed by Lambda) and knowledge bases, and Bedrock runs the reason–act loop for you. Bedrock AgentCore is the runtime for agents you built yourself — any framework (Strands, LangGraph, CrewAI, OpenAI Agents SDK), any model — deployed as containers with session isolation, memory, identity, and observability provided as services.
| Bedrock Agents | Bedrock AgentCore | |
|---|---|---|
What it is | Declarative managed agent: AWS runs the orchestration loop | Serverless runtime + services for agents you wrote yourself |
Framework | None — instructions and attachments | Any: Strands, LangGraph, CrewAI, OpenAI Agents SDK… |
Tools | Action groups (OpenAPI/function schema + Lambda) | Whatever the framework calls — plus Gateway, which turns existing APIs and Lambda functions into MCP tools |
Services around it | Knowledge bases, Guardrails, memory, versions/aliases | Runtime (session isolation), Memory, Identity (inbound/outbound auth), Code Interpreter, Browser, Observability |
Protocols | Bedrock APIs (InvokeAgent) | MCP for tools, A2A for agent-to-agent — standard protocols, hosted |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.