Arcade AI · Agent harness
An agent system that builds a whole game from one prompt
I'm a core engineer on the agent system that turns a written prompt into a playable 3D game, made reliable by its structure rather than by its prompt.
5 specialist sub-agents: world, logic, VFX, UI, audio
Read the details
The problem
A game is too big for one AI answer. It has a world, rules, effects, an interface and sound, and one mistake in any of them can break it. Asking the model to be careful does not make it reliable, and a run that goes wrong can keep spending without getting anywhere.
What I built
A harness, the structure the agents work inside: one tool-use loop (the model chooses a tool, uses it and reads the result), skills loaded only when they are needed, specialist sub-agents for the world, the game logic, the visual effects, the interface and the audio, and a memory kept for each level.
How I checked it
The checks are part of the system. Every output follows a typed contract, a strict format the code can verify. Validation gates reject an output that breaks its contract before it reaches the game. Cost ceilings cap what a run can spend, and failure breakers stop a run that keeps failing.
The result
Reliability that comes from the system, not from the prompt: a prompt goes in, and a playable 3D game comes out.
- specialist sub-agents: world, logic, VFX, UI, audio
- 5
- typed contracts, validation gates, cost ceilings, failure breakers
- 4 safeguards