The Problem
We kept coming back to one simple question:
Why does building a game have to start with knowing how to code?
AI can already generate code from a prompt, but there is still a big gap between AI-generated code and a playable game.
You still need to run it, debug it, modify it, and deploy it.
We wanted to make AI game development feel less like coding and more like having a conversation.
That became Lumio — a conversational AI game builder designed to turn natural-language ideas into playable browser games.
What Lumio Does
Lumio is an AI game generator where you describe the game you want in plain language.
For example:
“Create a game where a bee protects flowers from waves of enemies.”
The AI agent takes that prompt and works directly on the game.
It can read existing files, create new ones, modify code, and run the game using tools like read_file, write_file, replace_text, and list_files.
Once the game is built, it is packaged and delivered as a playable browser game.
And the conversation doesn't stop there.
“Add a boss.”
“Make the background darker.”
“Add sound effects.”
The AI agent works on the same game, so each prompt becomes another iteration.
That is the core idea behind Lumio:
Describe → Build → Play → Iterate
The AI Agent Behind Lumio
We use Amazon Bedrock for our AI models and Strands Agents for agent orchestration.
Game development isn't really a one-shot AI generation problem. The agent may need to understand the prompt, inspect the existing code, make multiple changes, run the game, and continue from the current state.
That's why we built Lumio around an AI coding agent rather than a simple prompt-and-response chatbot.
Strands Agents gives the agent the tools and workflow it needs to actually work on the game.
Where Does the Generated Code Run?
Once an AI agent can generate and execute code, we need to think about isolated execution.
We didn't want generated game code running directly inside our main application environment.
So each game gets an isolated workspace using Firecracker microVMs.
The agent can work with the game's source code, assets, runtime, and build process inside that environment.
The basic flow is:
User → AI Agent → Firecracker → Game
This gives our AI game builder a dedicated environment for creating and running generated games.
Getting the Game to the Browser
Building the game is only half the job. We also need to deploy and deliver it.
We use Amazon S3 to store generated game bundles and Amazon CloudFront to deliver them to the browser.
The flow is simple:
Game → Amazon S3 → Amazon CloudFront → Browser
This turns AI-generated game code into an actual playable web experience.
Keeping the Conversation Connected
A conversational AI game builder also needs persistent memory.
Lumio uses Amazon Aurora PostgreSQL to store game state and conversation history, along with references to the game bundle and sandbox.
So when a player says:
“Make the boss harder.”
Lumio knows which game they're talking about and can continue working from the existing version.
The AI conversation and game state stay connected instead of starting from scratch with every prompt.
What We Learned
The biggest lesson for us was that building an AI agent application is much more than choosing a model.
Once an AI needs to actually do something, you have to think about tools, execution, isolation, persistence, and deployment.
For Lumio, those pieces came together as:
Amazon Bedrock — AI models
Strands Agents — AI agent orchestration
Firecracker — isolated game execution
Aurora PostgreSQL — persistent game and conversation state
Amazon S3 — game bundle storage
Amazon CloudFront — game delivery
The interesting part isn't any one AWS service.
It's how they work together to create a complete AI-powered game development workflow.
Final Thought
We started with a simple idea:
What if you could build a game just by describing it?
That became Lumio.
You describe the game. The AI agent builds it. You play it.
Then you tell it what to change. And you keep going.
Lumio - If you can describe it, you can play it.
Loading article