My little AI-powered weekend game project is turning into something much bigger. What started as a WebGL experiment has now evolved into a Unity project with AI-generated 3D assets, multiple coding agents, and a growing list of technical challenges. And the development process has been anything but smooth. Here's what happened. FROM WEBGL TO UNITY The original project was built for the browser using WebGL. I decided to port it to Unity and continue developing it as a proper 3D game prototype. The first attempt was with Gemini 4 Hydrogen in Google Antigravity. Unfortunately, Hydrogen struggled with the transition from WebGL to Unity. Even with the relevant documentation available, it repeatedly misunderstood fundamental aspects of the Unity implementation. I tried getting help from Gemini 4 Argon, but that didn't resolve the problems either. Eventually, Argon exhausted its usage limits and was out of the game. ENTER GPT-6.1 TERRA I switched the project to GPT-6.1 Terra. And this time, things started moving forward. Terra managed to make substantial progress on the Unity port while continuing to develop and improve the project. But here's the catch. This isn't just a straightforward port anymore. New features, performance issues, scene organization, object rotations, and unexpected engine-specific problems keep appearing as development continues. AI-GENERATED 3D ASSETS Meanwhile, I also started building a proper asset pipeline. First, I generated 2D images with texture layouts. Then ChatGPT connected to fal.ai and used AI-powered 3D generation to turn those images into actual 3D objects. I delegated the next stages of asset production and integration to the agent as well. So now the workflow involves: • AI-generated concept art and texture layouts. • Image-to-3D asset generation through fal.ai. • WebGL-to-Unity migration. • AI-assisted scene construction and development. • Performance optimization and debugging. • Fixing object transformations and rotations in 3D space. • Multiple AI coding models taking over different stages of the project. THE REALITY OF AI GAME DEVELOPMENT The interesting part isn't that AI can generate code or 3D models. It's what happens when a project becomes complex enough that the agent has to understand the entire architecture, engine-specific behavior, dependencies, and performance constraints. Some models handle certain tasks surprisingly well. Others get stuck on fundamental issues, even when you provide documentation and additional guidance. And switching to a stronger model doesn't automatically eliminate every problem. That's the reality of building something more substantial than a one-prompt demo. The Unity version is still being refined, and I'll publish the updated online build once the remaining issues are resolved. For now, here's a small demonstration of the progress. Tools used: Unity Google Antigravity Gemini 4 Hydrogen Gemini 4 Argon GPT-6.1 Terra ChatGPT fal.ai How do you handle AI coding agents when one model gets completely stuck? Keep debugging or switch models? #Unity3D #AIGameDevelopment #VibeCoding