Max Chui · AWS AI League Hong Kong Champion

On 17 June 2026 I won the AWS AI League at AWS Summit Hong Kong, held at the Hong Kong Convention and Exhibition Centre. It was the first time the tournament had run in the city. The win came with a fully sponsored place at the Global Championship Grand Finale at AWS re:Invent 2026 in Las Vegas, where I’ll be representing Hong Kong.

The AI League is AWS’s competitive AI programme, run as a gamified tournament with live leaderboards, timed challenges, and real-time scoring. It offers two tracks: a Model Customization Challenge, where you fine-tune foundation models for a specific domain with Amazon SageMaker AI, and an Agentic AI Challenge, where you build an autonomous agent with Amazon Bedrock AgentCore and drop it into a maze-style environment that demands multi-step reasoning.

I entered the Agentic AI Challenge. More than 200 participants filled the Arena, including enterprise teams from Cathay, Manulife, Octopus, Crypto.com, and Lalamove, alongside individual developers. The build window was three hours.

Before I could build anything, the venue WiFi ground to a halt — the AWS Builder platform wouldn’t even load — so I switched to a mobile hotspot and burned at least ten minutes getting online. The system that followed was a multi-agent orchestration built from scratch inside what was left of the time limit. At its centre sat a supervisor agent on Bedrock AgentCore — a “Dungeon Game Orchestrator” — that routed each in-game encounter to the right resource: a pathfinding Lambda for movement decisions, specialist sub-agents for harder challenges, and a memory system that tracked challenge types and routing history so the same mistake wasn’t repeated.

Three decisions did most of the work. The first was prompt structure: explicit DO / DO NOT / MUST rules, plus a “no human is listening” framing that stopped the agent asking clarifying questions, since every question cost a life.

The second was replacing unreliable model guesses with deterministic tools. A Python-execution Lambda handled arithmetic an LLM would almost certainly hallucinate, such as a large Fibonacci term under a modulus, and a WebResearcher sub-agent fetched live pages for research questions instead of relying on recall.

The third was treating guardrails as a scoring lever rather than a safety afterthought. One challenge was won not by answering an adversarial prompt but by refusing it outright, which mirrors how responsible agents get designed in production.

It was close. With eleven minutes left I wasn’t on the first page of the leaderboard, and I was still optimising into the final minute. When the scores settled, Chu Ruisong — Corporate VP & President of AWS Greater China — came on stage to present the boarding pass and championship belt, and we had a short chat before the photos.

The Global Championship Grand Finale runs from 30 November to 4 December 2026 at AWS re:Invent in Las Vegas, against champions from AWS Summits worldwide, with finalists demonstrating their agents live and a US$50,000 prize pool on the table. You can follow along on my now page.