AgentCore: runtime, memory, gateway, identity

Lesson 3 of 4 in Agentic Systems on AWS.

Everything so far ran "somewhere". Amazon Bedrock AgentCore is AWS's answer to where: an agentic platform for building, deploying, and operating agents at scale with any framework and any model — Strands, LangGraph, CrewAI, LlamaIndex agents all deploy unchanged, calling models inside or outside Bedrock. Its services are modular (use them together or independently), and they map one-to-one onto the problems the previous lessons raised: a place to run the loop, a place to remember, a front door for tools, and credentials for all of it. The associate-level tour (Deploying FMs and agents) introduced AgentCore as a product; the professional exam expects you to know which service solves which problem.

Watch one user turn travel the platform, then read the service-by-service reference below the board.

Runtime — where the loop runs

Secure, serverless execution purpose-built for agents and tools: fast cold starts for interactive turns, extended runtime support for asynchronous agents, per-session isolation, and support for multi-modal and multi-agent workloads. Framework-agnostic (CrewAI, LangGraph, LlamaIndex, Strands, OpenAI Agents SDK) and model-agnostic (FMs in or outside Bedrock), with MCP and A2A protocol support.

Memory — what the agent remembers

Short-term memory for multi-turn conversations; long-term memory that persists across sessions, can be shared across agents, and lets agents learn from experience. You control what is remembered — memory is a managed store, not a side effect of the context window.

Gateway — tools without servers

Converts APIs, Lambda functions, and existing services into MCP-compatible tools, and connects to pre-existing MCP servers, exposing all of them to agents through gateway endpoints. The fastest path from "we have an API" to "the agent has a tool".

Identity — who the agent is

Agent identity, access, and authentication management, compatible with existing identity providers (Amazon Cognito, Okta, Microsoft Entra ID, Auth0) — covering inbound auth (who may invoke the agent) and outbound auth (what the agent presents to tools), with no user migration.

Code Interpreter and Browser — built-in tools

Code Interpreter: an isolated sandbox where agents execute code (Python, JavaScript, TypeScript) for calculation-heavy and data tasks. Browser: a managed cloud browser runtime so agents can navigate sites, fill forms, and extract information — works with automation frameworks like Playwright.

Observability — proof of what happened

Unified tracing, debugging, and monitoring: step-by-step visualizations of each agent workflow, inspectable execution paths and intermediate outputs, emitted in OpenTelemetry-compatible format so it plugs into your existing observability stack and CloudWatch.

Key terms: Bedrock AgentCore, MCP, Bedrock AgentCore Observability, Strands Agents

Interactive checkpoint quiz (3 questions) — open this page in a browser to take it.