Five services, one question each: where it runs, what state it keeps, what it can reach, what identity it acts under, and what it recorded. The catalogue is larger than five. What a production agent runtime has to provide goes deeper than this level needs, as do the two component cases: describing tools well enough to be chosen well and credentials scoped to the caller. At this level, remembering the five and what each one replaces is enough.
Flash card
Amazon Bedrock AgentCore: modular services for running an agent in production on AWS, used together or separately, with any framework and any foundation model. Adopting it is a hosting decision rather than a rewrite. Five services carry most of the ground at this level; others include Harness, Code Interpreter, Browser, Policy and Registry.
- Runtime is serverless and hosts an agent whatever framework wrote it, LangGraph, CrewAI and Strands Agents among them, or no framework at all. Each session gets its own microVM with isolated CPU, memory and filesystem, terminated and sanitised when the session stops. Sessions run up to eight hours by default, and context carries across invocations sharing a session ID.
- Memory holds short-term raw events for the session in progress, retained for up to 365 days, and long-term memories extracted across sessions by strategy: summaries, user preferences, semantic facts. Namespace templates keyed on actor and session scope what is kept per whom.
- Gateway converts APIs, Lambda functions and existing services into Model Context Protocol [MCP] tools, and fronts MCP servers already running. Semantic search across the catalogue narrows which tools reach the prompt; which one the model calls still depends on how each is described.
- Identity works in both directions. Inbound, a JWT authorizer validates who invoked the agent. Outbound, a token vault stores OAuth tokens, API keys and certificates and vends them per call, including on behalf of the signed-in user, so the agent never holds the long-term secret or refresh token.
- Observability emits OpenTelemetry-format traces of a run: which tools ran, in what order, with what inputs and returns, plus latency, token usage and error rates. It lands in Amazon CloudWatch, where a built-in dashboard renders runtime traces.
Pick it when
Pick it when an agent that runs on a laptop needs real users, and the choice is building the runtime, memory, tool plumbing and credential handling yourself.
It's the wrong answer when
It is the wrong answer for a fixed sequence of steps that never varies: that is orchestration of a known path, and AWS Step Functions runs it at lower cost with no model in the loop. A single prompt-and-response call with no tools is an Amazon Bedrock model invocation and nothing more.