Amazon Bedrock AgentCore Documentation
by Amazon Web Services
The reference for running agents in production on AWS: serverless runtime, memory, MCP gateway, identity, policy and evals as thirteen services you can adopt one at a time.
Overview
Amazon Bedrock AgentCore is AWS's managed platform for running agents in production, and its developer guide is the reference for all thirteen of its modular services, each usable independently of the others and of any particular framework. Runtime is a serverless environment with fast cold starts, true per-session isolation and support for long-running asynchronous agents. Harness is a higher-level managed agent loop invoked with a single API call, running each session in an isolated microVM with filesystem and shell access. Memory covers short-term multi-turn context and long-term stores that persist across sessions and can be shared between agents. Gateway converts existing REST APIs, Lambda functions and OpenAPI specs into MCP-compatible tools, and connects to pre-existing MCP servers. Identity brokers agent authentication against Cognito, Okta, Microsoft Entra ID or Auth0. Code Interpreter and Browser give agents sandboxed Python and JavaScript execution and a managed browser compatible with Playwright and Browser Use. Observability emits OpenTelemetry-compatible traces, while Evaluations, Optimization, Policy, Registry and Payments cover agent scoring over sessions and spans, A/B-tested prompt and tool-description tuning, Cedar-based deterministic guardrails on every tool call, an organization-wide catalog of agents and MCP servers, and x402 microtransactions for paid APIs. The guide is framework-agnostic throughout, with worked paths for CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands Agents, and model-agnostic across Bedrock, OpenAI, Gemini, Claude, Nova, Llama and Mistral. Pricing is consumption-based with no minimums; new AWS accounts get up to $200 in free-tier credits.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Documentation
- Cost
- Freemium
- Duration
- Reference documentation; ~2-3 hours to read the overview and deploy a first agent
- Provider
- Amazon Web Services
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Deploy an agent to a serverless runtime with true per-session isolation and fast cold starts
- ✓Give agents short-term conversational memory and long-term stores that persist across sessions
- ✓Turn existing REST APIs, Lambda functions and OpenAPI specs into MCP tools through Gateway
- ✓Attach agent identity and brokered OAuth credentials using Cognito, Okta, Entra ID or Auth0
- ✓Trace, debug and monitor production agent runs via OpenTelemetry-compatible observability
- ✓Sandbox untrusted code execution and browser automation with Code Interpreter and Browser
- ✓Enforce deterministic guardrails on every tool call by authoring Cedar policy rules
- ✓Score agent quality over sessions, traces and spans with the built-in Evaluations service
Highlights
- •Framework-agnostic and model-agnostic by design — CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands all documented, against Bedrock, OpenAI, Gemini, Claude, Nova, Llama or Mistral
- •Thirteen services that genuinely work standalone: you can adopt Memory or Gateway without moving your agent onto the runtime
- •Companion awslabs/agentcore-samples repo (3.3k stars, 650 commits) ships CloudFormation, CDK and Terraform templates plus end-to-end blueprints
- •Documents ground almost nobody else covers — x402 agent payments, a governed agent/MCP Registry, and A/B-tested prompt and tool-description optimization
- •Policy uses Cedar to intercept tool calls deterministically, which is a real answer to prompt-injection-driven tool abuse rather than another system-prompt plea
Who It’s For
Best For
- ✓Platform engineers building a paved path so product teams can ship agents on shared infrastructure
- ✓Teams already on AWS moving an agent prototype into production with identity, audit and observability requirements
- ✓Architects comparing managed agent runtimes against self-hosting on Kubernetes
Prerequisites
- •A working AWS account and comfort with IAM roles, policies and the AWS CLI
- •Experience building at least one agent in a framework such as LangGraph, CrewAI or Strands
- •Python or TypeScript, plus a working understanding of MCP if you plan to use Gateway
FAQ
What is Amazon Bedrock AgentCore Documentation?
The official developer guide for AWS's managed agent platform, aimed at engineers moving an agent prototype into production. It documents thirteen modular services — runtime, memory, an MCP gateway, identity, sandboxed code and browser tools, observability, evaluations, policy and more — each usable independently, with any open-source agent framework and any foundation model, so you are not buying a full-stack lock-in to get a runtime.
Is Amazon Bedrock AgentCore Documentation free?
Amazon Bedrock AgentCore Documentation offers free content, with paid options for certificates or premium features.
What level is Amazon Bedrock AgentCore Documentation for?
Amazon Bedrock AgentCore Documentation is aimed at a intermediate audience. Recommended background: A working AWS account and comfort with IAM roles, policies and the AWS CLI, Experience building at least one agent in a framework such as LangGraph, CrewAI or Strands, Python or TypeScript, plus a working understanding of MCP if you plan to use Gateway.
How long does Amazon Bedrock AgentCore Documentation take?
Expect roughly Reference documentation; ~2-3 hours to read the overview and deploy a first agent. Most learners work through it at their own pace.
What will I learn from Amazon Bedrock AgentCore Documentation?
You'll learn: Deploy an agent to a serverless runtime with true per-session isolation and fast cold starts; Give agents short-term conversational memory and long-term stores that persist across sessions; Turn existing REST APIs, Lambda functions and OpenAPI specs into MCP tools through Gateway; Attach agent identity and brokered OAuth credentials using Cognito, Okta, Entra ID or Auth0; Trace, debug and monitor production agent runs via OpenTelemetry-compatible observability; Sandbox untrusted code execution and browser automation with Code Interpreter and Browser; Enforce deterministic guardrails on every tool call by authoring Cedar policy rules; Score agent quality over sessions, traces and spans with the built-in Evaluations service.
Topics
Sources
This page was written from 3 sources, 2 on domains other than docs.aws.amazon.com.