AWS's agent hosting layer: bring your own framework, pay for the infrastructure

What it does

This repository contains example code and tutorials for Amazon Bedrock AgentCore, AWS’s managed infrastructure for deploying AI agents. The service provides a serverless runtime, identity management, memory, observability, and policy enforcement—essentially all the unglamorous plumbing required to move an agent from a laptop prototype to something that won’t wake someone up at 3 AM.

The interesting bit

AgentCore is deliberately framework-agnostic and model-agnostic. You can build with LangGraph, CrewAI, LlamaIndex, Strands Agents, or “any other framework,” then deploy through AWS without rewriting. The Gateway feature converts existing APIs and Lambda functions into MCP-compatible tools, which suggests AWS recognizes most enterprises already have services they want agents to call rather than building fresh tool ecosystems.

Key highlights

• Serverless runtime for agents and tools with OpenTelemetry-based observability
• Managed memory infrastructure for persistent agent sessions
• Cedar policy engine for fine-grained access control
• Built-in evaluation framework including LLM-as-a-Judge and continuous online evaluation
• Infrastructure-as-code templates for CloudFormation, CDK, and Terraform
• CLI (agentcore create, agentcore deploy) abstracts AWS provisioning
• Identity provider integrations: Okta, Entra, Cognito

Caveats

• Repository is actively migrating from “Starter Toolkit” to new AgentCore CLI; legacy samples remain in legacy/ and are being updated “over the coming weeks”
• Requires BedrockAgentCoreFullAccess and AmazonBedrockFullAccess managed policies—this is not a lightweight permission footprint
• Some samples are Jupyter notebooks, which is an odd format for production deployment examples

Verdict

Worth exploring if you’re already committed to AWS and need to productionize agents built with popular frameworks. Skip it if you’re looking for a quick local experiment or if multi-cloud portability is a priority—the lock-in is structural, not incidental.