AWS Builder Center
A Builder’s Guide to AWS GenAI and Agentic AI Repositories 

A Builder’s Guide to AWS GenAI and Agentic AI Repositories 

This article helps you find the right AWS repositories for your GenAI and agentic AI journey. You’ll explore samples for learning Amazon Bedrock, building agents with Strands, operating them with Amazon Bedrock AgentCore, connecting tools through MCP, and implementing advanced RAG and GraphRAG patterns. By the end, you’ll know which repositories to start with based on what you want to learn or build.

Technical Program Manager, AI/ML

If you’ve been trying to keep up with generative AI lately, you know the feeling: new agent frameworks, MCP servers, reference architectures, and samples seem to appear every week. Great for builders. Slightly less great when you are staring at hundreds of repositories wondering: where do I start? 
AWS already has hundreds of actively maintained samples on GitHub. They cover everything from GenAI foundations to agents, memory, RAG, tool use, and production-ready patterns.  The challenge with having many high-quality options to choose from is finding the right one for where you are on your journey. I curated samples for each stage of the GenAI and agentic AI learning journey.  
Find your stage and start there. 

Stage 1: Learn the Foundations of GenAI on AWS 

 amazon-bedrock-samples 

The canonical “start here” repository. Pre-built notebooks covering the full breadth of Amazon Bedrock: prompt engineering, agents and function calling, RAG, multimodal, custom model import, responsible AI, and, critically, a POC-to-prod section for when your experiment needs to become real. There’s also a browsable  companion website . 
Start with: the introduction-to-bedrock folder, then pick the use-case folder that matches your project. 

 amazon-bedrock-for-beginners 

A step-by-step walk-through for developers whose main question is "how do I call a model from my code?" It uses the raw Bedrock APIs first, building up one concept at a time: a first Converse API call, tokens and cost, multi-turn conversations, tool use, RAG with Knowledge Bases, and Guardrails. It ends by tying everything into a university FAQ chatbot, then rebuilding it with the Strands Agents SDK so you see what the framework handles once you know the basics. Comes with a  companion repo  and an  accompanying video .
Start with: the Converse API example, then work through the sections in order.

 amazon-nova-samples 

Working examples for Amazon’s own Nova model family, worth knowing because it covers ground the others don’t: image generation (Canvas), video generation (Reel), and low-cost text models for high-volume workloads.

Stage 2: Build Your First Agent with Strands

 Strands  is an open source toolkit for building production agents. It includes the Strands Harness SDK, an open source agent-building framework that lets the model drive the control flow while the SDK handles the underlying loop mechanics.

 strands-agents/harness-sdk 

An SDK for building agents in Python and TypeScript, with lifecycle controls, tools and structured output, MCP support, multi-agent patterns, memory and sessions, streaming, guardrails, tracing, and evaluations. It’s model-agnostic, with support for Amazon Bedrock, Anthropic, OpenAI, Gemini, and more, so your agent code can survive a model switch. 
Your first agent can use tools shipped directly with the Strands SDK: 
from strands import Agent
from strands.vended_tools import file_editor
agent = Agent(tools=[file_editor])
agent("Create a file at /tmp/hello.txt containing 'Hello from Strands!'")
Four lines: instantiate an agent, register the SDK’s vended file editor, and give it a task. The model decides to call the tool and supplies the required arguments. That’s the agent loop in miniature. 

 Strands Agents learning course

A structured, self-paced course for learning agent development with Strands. A good fit if you prefer a guided path over exploring samples. Follow the course on  StrandsAgents.com  along with the companion repo or watch the  14-video series on AWS Builder Center.  

 strands-agents/samples 

Ready-to-run examples built on the SDK, and the fastest way to see agent patterns (tool use, multi-agent, streaming) in working code. Note these are educational samples, not production blueprints; that’s what Stage 3 is for.

 strands-agents/evals 

An evaluation framework for agents and LLM applications. If you come from data science, you already know models don’t ship without evaluation; agents are no different, yet it’s the most-skipped step in the field. Bookmark this before you need it.

 strands-labs 

The experimental arm of Strands, hosting early projects in areas such as benchmark harnesses, harness optimization, AI functions, and physical AI. Explore it when you want to test emerging ideas beyond the core toolkit.

Stage 3: Take Agents to Production with Amazon Bedrock AgentCore

Running agents in production raises problems that never come up in a prototype: session isolation, persistent memory across conversations, authenticating the agent to other systems, observability into what it actually did, and scaling.  Amazon Bedrock AgentCore  is AWS’s managed infrastructure for that layer, and it’s framework-agnostic and model-agnostic: bring Strands, LangGraph, CrewAI, LlamaIndex, or your own loop.

 awslabs/agentcore-samples 

The main AgentCore samples repository, organized exactly how you’d want to learn:
  • getting-started/: your first agent in minutes with the new  AgentCore CLI , in Python or TypeScript
  • features/: one folder per capability, covering Runtime, Gateway (turn any API or Lambda into MCP tools), Identity, Memory, built-in tools (Code Interpreter, Browser, Web Search), Observability, Evaluation, and policy-based access control
  • end-to-end/: complete production-ready applications with architecture diagrams
  • infrastructure-as-code/: CloudFormation, CDK, and Terraform templates
The repo is mid-migration from the older Starter Toolkit to the AgentCore CLI. Start with the new getting-started/ paths, not legacy/.

 aws/bedrock-agentcore-sdk-python 

The Python SDK for adapting an existing agent to AgentCore Runtime. Wrap an agent from any framework in a BedrockAgentCoreApp entry point.

 aws/agentcore-cli 

The recommended command-line tool for scaffolding, developing, deploying, and invoking agents on Amazon Bedrock AgentCore. Use it to take your SDK-wrapped agent from local development to AgentCore Runtime.

Stage 4: Extend with MCP Servers and Agent Tooling

 awslabs/mcp 

The most popular repo in this entire list: AWS’s suite of open source MCP servers, covering documentation lookup, infrastructure-as-code, containers, serverless, databases, cost analysis, and much more. If you use Claude, Kiro, Cursor, or any MCP-capable assistant, these servers make it fluent in AWS.
AWS has announced the  Agent Toolkit for AWS  as the successor to these MCP servers for production use. It adds IAM condition keys to distinguish agent actions from human ones, plus CloudWatch and CloudTrail visibility. The awslabs/mcp repo remains active, but if you’re building for production, check out the Agent Toolkit first.

 kirodotdev/KiroCrew 

An open source, persistent development workspace powered by Kiro CLI. It preserves context and learning across sessions and can run multi-step tasks, scheduled jobs, and monitoring workflows unattended.

 Agent Toolkit for AWS 

For production agent tooling on AWS, start with the Agent Toolkit for AWS. It adds AWS-aware authorization and observability, including IAM controls that distinguish agent actions from human actions and visibility through CloudWatch and CloudTrail.

Stage 5: Go Deeper with RAG, Graphs, and Multi-Framework Patterns

 awslabs/graphrag-toolkit 

Python tools for graph-enhanced GenAI. Its lexical-graph framework automatically builds a hierarchical knowledge graph from unstructured data and composes question-answering strategies over it. Where vector RAG retrieves passages by embedding similarity, GraphRAG also captures the relationships between entities across documents, which pays off on multi-hop questions that similarity search alone gets wrong. Includes a self-guided workshop with Amazon Neptune.

 aws-samples/amazon-bedrock-rag 

Fully managed RAG using Bedrock Knowledge Bases, which handles the chunking, embedding, and vector storage for you. The simplest credible RAG architecture on AWS.

 awslabs/generative-ai-cdk-constructs 

Well-architected, multi-service CDK patterns for GenAI infrastructure in TypeScript or Python, so the infrastructure behind everything above is repeatable code, not console clicks.

 aws-samples/sample-agentic-frameworks-on-aws 

If you’re already invested in LangGraph, CrewAI, LlamaIndex, or mem0, this repo shows how to run those frameworks on AWS: reference architectures, notebooks, linked workshops, and companion blog posts covering memory, A2A protocol, multi-agent collaboration, and observability with Langfuse.

Where to Start (If You Only Pick Three) 

AWS’s GenAI and agentic AI story follows a clear path: Bedrock for models, Strands for building agents, AgentCore for operating them, and MCP for connecting them to everything else. 
If you’re already midway through your journey, skip ahead. Pick the stage that matches your current challenge, clone one repository, and build something small with it this week. 
Then share what you built, what you learned, or which repository you think belongs in the next version of this guide. Which one will you start with? 
Happy building! 🚀
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article