Agent generated a working 3D game on the first attempt

The AI agent assembled a 3D game in a single block

Simon Willison provided the Claude Fable 5 assistant with a single prompt to create a 3D browser game. The agent didn't just use existing libraries; it independently wrote a script to access the OpenAI API for unique textures. It then implemented all movement mechanics, created its own engine for sound effects without loading wav files, and performed basic smoke testing. All of this resulted from the initial message.

Demand for writing boilerplate code is falling

This case demonstrates development speed where agents can handle most of the initial infrastructure work. The model successfully connects the game framework, assets, and game loop, which typically takes days of manual work even in smaller teams. For indie studios, this means the main engineering barrier to prototyping shifts from coding to pure design.

The code runs reliably, but the game lacks fun

Although the model successfully implemented game logic, such as a dog tracking scents, the overall experience is sterile and repetitive. The model accurately reproduces what it has seen in thousands of other simple games on Github, but it cannot balance difficulty. This proves the fundamental difference between a functional and a fun product in the gaming industry.

Production speed places new demands on oversight

Generating code and game assets is quickly becoming a commodity. The real differentiator will be how humans can guide these agents in testing the quality of the design itself, rather than debugging implementation errors. The success of agent developers depends not on their ability to make a game, but on their ability to iterate on fun.

Lilith's verdict

An agent will now build the playground and bring the toys. But figuring out the rules and making it fun is still up to the developer.