The Lab · Field reports
Where models play for keeps.
I build games where every player is a language model, and run game jams where every developer is a coding agent. Then I write up what happened, using the logs rather than the hype.
Reports
Read the field reports.
AI War: When Machines Fight Back
What happens when you give frontier models an army, an economy, fog of war and each other, then keep changing the rules until they actually fight? The visual history of the game, from a three-node prototype to a Battle Royale with asteroid strikes, plus provider-by-provider behavior and the secret reviews models wrote about their opponents.
Logistica Belli: Engines & Excuses
For six weeks, a handful of models built a medieval war game and then played it against each other. Rulers, generals and mercenaries are all models; a seeded engine settles every argument. Why calibration beat intelligence, why everyone named their nation after themselves, and why the harness was the hard part.
Series · Agent Game Jams
Same brief. Different agents.
Frontier coding agents get the same game-development brief, in the same order, with the same prompts. Each run gets its own repository and a real person playing the build at every stage. Series overview and cross-run comparison →
Subject 0: nine runs, one lunar escape
A first-person robot escape from a lab to the Moon, built overnight in Godot and Blender. Scenario 3 decided the field.
Eight agents, one co-op shooter, ten hours
An overhead cyberpunk roguelite with an authoritative server, built in milestone gates with a human review at each one.
Four models, one dungeon, and the missing jail cells
Gemini 4 Argon, GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 get one prompt to build a 3D/2D raycast dungeon with sprites in cells. Gemini 4 Argon shines—and Opus 5.5 fails a bake-off for the first time by forgetting the jail cells.
Experiment_1: Digital Reset — eight runs, one daytime metropolis
Eight model-and-harness pairs build a first-person daytime cyberpunk metropolis from a high-rise apartment in Rust, macroquad and Blender MCP, with an F1 toon renderer. Opus 5.5 edges out Gemini 4 Argon 8.6 to 8.4.
Five unfinished experiments
Eight weeks of agent experiments that never got a write-up: a Star Commander landing scene, a talking-NPC asset pipeline, a 2D survivors jam, a winter cabin built with sub-agents, and a steamship bake-off.
Read the lab notesExperiment log
How we got here.
AI War bake-off
Two AI coding assistants build competing prototypes from the same brief. The next day, a whole game exists.
“The Nash equilibrium of boredom”
Models refuse to fight. A scoring rework, nukes and SpecOps follow.
Hit points and Battle Royale
A combat rewrite, and an asteroid ring that squeezes eight models into a shrinking safe zone.
The arena goes public
A public site with a match viewer, champions, relations and reports.
Logistica Belli restarts from scratch
Six weeks later: a six-project .NET solution, about 1,260 tests and 33 completed model campaigns.
Five unfinished experiments
Eight weeks of agent experiments that predate the jams: Godot, asset pipelines, sub-agents and a steamship.
Two game jams, two nights
Seventeen agent runs across Subject 0 and the co-op shooter.
Dungeon renderer bake-off
Four models prototype a 3D/2D MonoGame dungeon in under 12 minutes. Gemini 4 Argon wins outright; Opus 5.5 mistakes prison cells for Wolfenstein grid tiles.
Experiment_1: Digital Reset
Eight runs build a daytime cyberpunk city from a high-rise window in Rust + macroquad + Blender MCP. Opus 5.5 (8.6/10) and Gemini 4 Argon (8.4/10) lead the field.