Amazon Bedrock AgentCore Samples

Overview#

Amazon Bedrock AgentCore Samples is the official practice repository maintained by AWS Labs (awslabs), designed to help developers master and implement the core capabilities of the Amazon Bedrock AgentCore platform. The platform adopts a framework-agnostic and model-agnostic design philosophy, natively compatible with mainstream orchestration frameworks including Strands Agents, CrewAI, LangGraph, LlamaIndex, Google ADK, and OpenAI, as well as any LLM.

Core Capabilities#

Agent Runtime & Deployment#

• Serverless Runtime: Pay-per-active-resource billing, supporting workloads from low-latency conversations to 8-hour asynchronous tasks with strict session isolation
• Multi-mode Deployment: Support for direct code upload or container image deployment to AgentCore Runtime

Tool & Protocol Integration#

• MCP-Compatible Gateway: Seamlessly convert existing APIs and Lambda functions into MCP (Model Context Protocol) compatible tools with semantic search-based tool discovery
• Code Interpreter: Multi-language code execution in secure sandbox environments
• Serverless Browser: Serverless browser runtime enabling agents to interact directly with web applications

Security & Governance#

• Identity Management: Support for agents to access AWS resources and third-party services on behalf of specific users or autonomously, with integration for enterprise IdPs including Okta, Microsoft Entra ID, and Amazon Cognito
• Policy Engine: Fine-grained access control based on the Cedar policy language, supporting natural language to policy translation and real-time interception of non-compliant tool calls

State Management & Quality Assurance#

• Managed Memory: Cross-interaction context maintenance infrastructure supporting personalized agent experiences
• Evaluation System: Built-in and custom evaluators supporting on-demand and online continuous evaluation
• Full-chain Observability: Based on the OpenTelemetry standard, natively integrated with Amazon CloudWatch and third-party monitoring tools including Grafana, Datadog, and Dynatrace

Architecture Highlights#

• Progressive Directory Structure: Repository organized by numbered progression—00-getting-started → 01-tutorials → 02-use-cases → 03-integrations → 04-infrastructure-as-code → 05-blueprints—with complete architecture diagrams and test instructions
• MCP Protocol Translation Layer: MCP standard adaptation at the gateway layer with semantic search engine for dynamic tool routing and discovery
• Cedar Policy Interception: Policy engine as tool call middleware leveraging Cedar syntax parsing for sub-second security interception
• OpenTelemetry Telemetry Pipeline: All component Traces, Metrics, and Logs exported following OpenTelemetry specifications
• Enterprise Network Isolation: Underlying architecture supports VPC connectivity and AWS PrivateLink

Quick Start#

Prerequisites:

• AWS account with configured credentials (aws configure)
• Node.js 20.x or higher
• uv (for Python agents) or Node.js (for TypeScript agents)
• Anthropic Claude 4.0 model access enabled in Amazon Bedrock console
• IAM permissions: BedrockAgentCoreFullAccess and AmazonBedrockFullAccess managed policies

npm install -g @aws/agentcore
agentcore create
cd my-agent
agentcore dev
agentcore deploy
agentcore invoke

Ecosystem Integration#

• Orchestration Frameworks: Strands Agents, CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI
• Identity Providers: Okta, Microsoft Entra ID, Amazon Cognito
• Observability Platforms: Amazon CloudWatch, Grafana, Datadog, Dynatrace
• Frontend Patterns: Streamlit, AG-UI
• Infrastructure as Code: AWS CloudFormation, AWS CDK, Terraform

Pending Confirmations#

• AgentCore GA date not explicitly stated
• Browser Tool regional availability unspecified
• Discord community full URL not resolved from page text
• Python SDK specific PyPI package name to be confirmed
• Starter Toolkit to AgentCore CLI migration timeline stated as "coming weeks" without specific date