Open Source: Practice the AWS AI League Agentic Challenge with the Community Edition

The AWS AI League agentic challenge is unlike anything the community has tackled before. You’re not just training a model - you’re building a multi-agent system with supervisor agents, sub-agents, tools, guardrails, and memory. If you’re having your first attempt at an AWS Summit, the learning curve is steep, competition time is limited. The only way to prepare or get better outside of the competition windows is to practice.

That’s why we’ve open-sourced, starting with two repositories under the aws-ai-community GitHub organisation to help the community prepare for events, as well as learn more about previous events.

AI League Community Edition

Repository: github.com/aws-ai-community/ai-league-community-edition

A full-stack practice platform that provides a good approximation of the real AWS AI League agentic challenge environment, for personal use. Deploy it to your own AWS account and practice building agents without waiting for the next competition window. This will be evolving over time, feel free to add issues to the GitHub repo if you find them, or for new things you’d like to see, or better still send us a PR!

What you get

• Map Builder - Create custom dungeon maps with a drag-and-drop tile editor. Set challenges, configure points/damage, and validate paths to treasure. Comes complete with ’tool tips’ to explain each challenge, along with the option to choose from a bank of questions from real AWS competitions, or LLM generate your own.

• Agent Builder - Configure your supervisor agent, attach sub-agents with their own prompts and models, register Lambda tools, create guardrails with content filters and deny topics, and attach memory for persistent recall.

• Gameplay - Run your agent against any map. Watch the animated replay with real-time combat log showing the full prompt, pathfinding response, challenge questions, and agent answers.

• Leaderboard - Track your scores across maps. Agent configurations are versioned on each submission.

Architecture highlights

The platform uses the same technologies as the real competition: -

graph TD
    Browser[Browser] --> CF[CloudFront + S3]
    Browser --> Cognito[Amazon Cognito]
    CF --> AppSync[AppSync GraphQL API]
    AppSync --> Lambda[Agentic API Lambda]
    Lambda --> DDB[(DynamoDB<br/>Maps · Sessions · Configs<br/>Leaderboard · Submissions)]
    Lambda --> GameRunner[Game Runner Lambda]
    GameRunner --> AgentCore[AgentCore Runtime<br/>Container]
    AgentCore --> Supervisor[Supervisor Agent<br/>Strands SDK]
    Supervisor -->|delegates| SubAgent[Pathfinder Sub-Agent]
    SubAgent --> Gateway[AgentCore Gateway<br/>MCP Protocol]
    Gateway --> PathfinderTool[Pathfinder Lambda<br/>BFS swift / get_coins]
    Lambda --> Bedrock[Amazon Bedrock<br/>Nova · Llama · Mistral · Claude]
    GameRunner --> AgentCore2[AgentCore Runtime<br/>Challenge Q&A]

    style AgentCore fill:#f9f,stroke:#333
    style Gateway fill:#bbf,stroke:#333
    style PathfinderTool fill:#bfb,stroke:#333

• Amazon Bedrock AgentCore Runtime — Your agent runs as a container with the Strands SDK.
• AgentCore Gateway — Lambda tools are exposed via MCP protocol through the Gateway, matching the competition setup.
• Supervisor → Sub-agent → Tool — The default configuration ships with a supervisor that delegates pathfinding to a sub-agent, which calls a BFS pathfinder Lambda tool. This mirrors the architecture you get at the start of the competition.
• Model flexibility — An enhancement over the AWS competition, you can choose from Amazon Nova (Micro, Lite, Pro), DeepSeek, Meta Llama, Mistral, or Anthropic Claude. Nova and the open weight models are typically covered by AWS credits, Claude models typically are not.

Deployment

git clone https://github.com/aws-ai-community/ai-league-community-edition.git
cd ai-league-community-edition
npm install
npx cdk bootstrap aws://<ACCOUNT_ID>/<REGION>
npx cdk deploy

CDK handles everything - Cognito auth, DynamoDB tables, AppSync API, CodeBuild for the agent container, AgentCore Runtime provisioning, CloudFront distribution, and admin account seeding. Once deployed it’ll share your URL and admin password in the CLI.

Why practice matters

During the competition you get limited time. The agentic challenge involves: -

• Getting your supervisor to correctly delegate to specialist sub-agents
• Ensuring your pathfinder tool receives the right parameters (map data, start position, strategy)
• Tuning prompts so models follow instructions precisely (especially with tool use)
• Configuring guardrails that actually block harmful content (empty responses don’t count!)
• Managing token usage across multiple agent invocations per game

All of these are things you can only learn by doing. The Community Edition gives you unlimited attempts to iterate.

Wealth Warning!!!

LLM usage can rack up large bills, which may not be covered by AWS credits, if your usual approach is to fund your AWS usage this way. You are responsible for any costs you incur in your AWS account. If you’re using AWS credits we’d recommend you try a model and wait to see whether it’s covered by credits by checking the billing console ~24 hours later to see if credits zero’d out your costs. Neither the AWS Community nor AWS themselves will help you with your bill. A full game run when you’re visiting 15-16 challenges could cost several dollars with the higher cost models.

No need to worry about the general hosting of the app when not in use - it’s all serverless and low cost, but you can of course destroy and recreate if you wish, just remember to delete the resources you create with the game, e.g. additional Lambdas, memory, guardrails, agents, as the CDK destroy won’t delete them.

AI League Competition Data

Repository: github.com/aws-ai-community/ai-league-competition-data

Historical maps and challenge definitions from past AWS AI League events. Use these to:

• Study map layouts - See what grid sizes, challenge placements, and time limits the real AWS competitions have used
• Test your agents against real maps - Load competition maps into the Community Edition Map Builder
• Analyse scoring strategies - Understand which maps you may want to optimise for score vs. survival

Available competitions at the time of writing

Each competition includes the map JSON, a rendered PNG visualisation, and a README with event context.

If you attend a future event please share the details with us, or better still, submit a PR.

Contributing

Both repositories welcome contributions: -

• Competition data — If you participated in an event and captured the map data, submit a PR with the JSON and any context you can share.
• Community Edition — Bug fixes, new challenge types, improved prompts, or new tool implementations. Check the issues and roadmap in the README.

What’s next

The Community Edition roadmap includes: -

• Lambda tool creation via SageMaker Code Editor
• Model Workshop for fine-tuning

As more AWS AI League events happen globally, we’ll continue adding competition data and improving the practice platform.

Get started: github.com/aws-ai-community/ai-league-community-edition

See you on the leaderboard.