AWS AI League 2026 Championship Kicks Off: Build Agents, Fine-Tune Models, Compete Globally
Compete at AWS Summits worldwide by building AI agents with Amazon Bedrock AgentCore and fine-tuning models on Amazon SageMaker AI. Top builders advance to the Grand Finale at re:Invent for a $50,000 prize pool
AWS AI League is a hands-on AI competition where developers, architects, and data scientists of all levels tackle real-world challenges using AWS AI services under a time constraint, on a live leaderboard. Starting April 22 at AWS Summit London & Bengaluru, the 2026 Championship season officially begins and it runs in person across 9 AWS Summit worldwide!
At each Summit, there are three ways to compete:
- Workshop: A deep, hands-on session where you build and optimize an AI agent from scratch. The top 3 on the leaderboard win $1,000 in AWS credits each and an invitation to compete in the Gameshow.
- Gameshow: The high-stakes showdown. The 1st place winner earns a ticket to re:Invent 2026 in Las Vegas, including travel and accommodations, plus qualification for the AI League 2026 Championship Grand Finale at re:Invent.
- Kiosk at the Summit Expo Floor: Drop by the AI League kiosk for a quick challenge. The top 100 participants earn $100 in AWS credits. Plus, everyone who joins the AWS AI League Space in Builder Centre and completes a survey receives $25 in AWS credits.
The Championship path is: compete at a Summit → win the Gameshow → advance to the Grand Finale at re:Invent in Las Vegas (Nov 30 – Dec 4, 2026) where the AWS AI League Champion will be crowned and a $50,000 prize pool is up for grabs.
Can't make it to a Summit in person? We're building a Virtual League on AWS Builder Centre so you can compete from anywhere. More on that later!
What you'll actually build with AWS AI League
The challenge: build intelligent agents that navigate a dynamic, maze-like environment where your agent encounters real-world challenges, from content moderation and code execution to browser interaction while racing to collect points and reach the objective. At the end of the competition you'll fine-tune a foundation model to sharpen your agent's reasoning, control its navigation and decision-making while optimizing for both accuracy and token efficiency.
Here's what that looks like in practice:
1. Build and orchestrate agents with Amazon Bedrock AgentCore
You'll configure a multi-agent system using Amazon Bedrock AgentCore - a supervisor agent that delegates to specialized sub-agents, each backed by AWS Lambda tools for pathfinding, web search, and code execution. You'll set up AgentCore Memory for contextual awareness across conversations, configure Amazon Bedrock Guardrails for safe agent behavior, and write or modify Lambda functions to extend your agent's capabilities. The agent runtime routes prompts between a default model on Amazon Bedrock and your fine-tuned model on a Amazon SageMaker AI endpoint.
2. Fine-tune on Amazon SageMaker AI
The new model customization experience in Amazon SageMaker AI gives you serverless model customization, an evaluation suite, experiment tracking, governance, 1-click deployment, and synthetic data generation. Models supported include text-to-text models from Nova, Qwen, gpt-oss, and Llama families. Techniques supported: SFT, GRPO with RLVR / RLAIF, and DPO, all in either full-rank or PEFT-LoRA. Flexible interfaces include an agentic model builder assistant, Studio point-and-click UI, and SDK.
The training workflow is two-step. First, you train a base model (like Qwen3 .6B) on prompt examples with a custom reward function to ensure correct tool calling. Then you take that tuned model and run a second pass focused on ensuring faithfulness. Each training run takes roughly 10–20 minutes (2 epochs). Once trained, you configure your agent in the agent builder, submit to the leaderboard, and see how you stack up.
3. RLVR: Training agents with measurable success
This is where it gets interesting for practitioners. The fine-tuning approach uses Reinforcement Learning with Verifiable Rewards (RLVR) - an objective, rule-based training approach using measurable rewards, not subjective human judgment.
Why this matters for agent development: tasks like tool calling, code execution, and navigation have clear, verifiable success criteria. The training loop works like this:
1. Model attempts task → Executes action in environment
2. System evaluates → Applies verifiable metrics
3. Model learns → Updates based on measurable feedback
4. Cycle repeats → Continuous improvement
2. System evaluates → Applies verifiable metrics
3. Model learns → Updates based on measurable feedback
4. Cycle repeats → Continuous improvement
In the AI League challenge, this translates to verifiable rewards like correct tool call responses and Lambda execution results, clear optimization where the model learns how to call tools effectively, immediate feedback with no waiting for human evaluation, and scalable learning with thousands of iterations possible.
The key theme: change the targeted behavior of a model rather than steer it at runtime or give it broad knowledge. Fine-tuning is the right approach when you need specialized task improvement, precise behavioral alignment, complex reasoning patterns, enforced format and structure, or optimized latency and cost at scale, exactly the kind of objectives you'll tackle in the competition.
4. Compete on the live leaderboard
Your agent is scored by an automated AI judging system based on accuracy, efficiency, task completion, and response quality. Scores are calculated automatically and reflected on the leaderboard in real-time - including your best score, total submissions, and token efficiency. You test on a game map, then submit to an evaluation map for your official leaderboard score.
How to Participate
There are two ways to join the AWS AI League Championship this year:
Option 1: Compete in person at an AWS Summit
The Championship kicks off on April 22 and continues across 9 Summits and it leads to re:Invent
| Date | Summit Location |
|---|---|
| April 22 | London, UK |
| April 22–23 | Bengaluru, India |
| May 14 | Sydney, Australia |
| May 20–21 | Seoul, South Korea |
| June 17 | New York City, USA |
| June 17 | Hong Kong |
| June 23–24 | Shanghai, China |
| June 30 – July 1 | Washington, D.C., USA |
| August 12 | Mexico City, Mexico |
Register for an upcoming AWS Summit at aws.amazon.com/events/summits and look for the AI League competition on-site. No purchase necessary. Must be 18+ to participate. See official contest rules for full eligibility details and restrictions.
Option 2: Join the Virtual League on AWS Builder Centre
Can't attend a Summit? The Virtual League is coming to AWS Builder Centre - the same challenges, the same leaderboard, accessible from anywhere, launch is planned for later this year.
Join the AWS AI League space on AWS Builder Centre now to be the first to know when the Virtual League opens for registration. You'll get updates on launch dates, challenge details, and early access opportunities.
Bring AI League to your organization
If you're a technical leader looking to upskill your team, you can also bring AWS AI League into your own organization. Customized challenges built around your industry, your people, your pace, with support from AWS teams. Reach out through the AI League website to learn more.
The bottom line
The 2026 Championship starts April 22 with A $50,000 prize pool and the title of AWS AI League Champion await at re:Invent. Come show us what you've got!
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article