My experiments with game-like interfaces for agents drove me to fulfill that high school dream of building my own videogame.
I picked Godot not wanting to get too heavily tied in the commercial world of Unreal engine or Unity.
Readers of this blog will not be surprised that all of my work on the game has been via agentic orchestration. For the last month or so a big portion of my token budget has been dedicated to building a game exploring maintaining the Hetch Hetchey pipeline system 20 years after the collapse of civilization.
One limitation I quickly hit with existing Godot tooling was the ability to have agents automatically play test and visually validate changes. There were a couple of not great MCPs for Godot focused on automating the Godot IDE, but nothing to do with automating the actual game.
Starting out with the idea of “I want a tool that can do for Godot games that can do everything Playwright does for the web” sounds absurd when you consider how many years it took to develop Playwright to where it is and the number of lackluster tools in that same vein produced over the last 20 years.
I built this tool because my ideal game dev workflow needed this tool to exist. No grand motivation, no desire to become a fulltime game programmer. Just a gap that I felt confident I could fix with modern AI tooling.
I’ve noticed this pattern a lot lately as I’ve fully internalized being on the other side of agentic development. Creating software is cheap, cheaper than it ever has been. The number of ideas that went from 3 month undifferentiated heavy lift nightmare to a quick weekend project keep increasing.
When your actual goal is to build a game you need some serious motivation to pivot and spend more effort on a tool supporting the game development process than the game itself. Now that decision becomes how much of my current 5 hour window do I dedicate to game tasks vs tooling tasks.
The game build itself is still grinding along, my ambition about two sizes too large for my non-existent game dev experience. Turns out visuals and playability are just as big a part of making a fun game as well written game code.
Godot Stagehand is usable with some patience. I run into frequent issues with agents fighting over sessions and persistence is a recurring problem. Shipping screenshots was surprisingly challenging and took some serious tweaking to get right.
The main takeaway is that building a sophisticated tool that serves no purpose than helping me with my main task was enough of an afterthought that I just built it. How many projects like this are currently sitting in your way at your workplace or on your side project that might be solved by building something new that you would have never considered building a year ago?