Building a Game with the Gauntlet Loop

The Gauntlet Loop Methodology

The Gauntlet Loop is a recursive prompting architecture that automates software development by delegating tasks to specialized sub-agents. The process begins with a high-level goal and a target reference. The system initializes a workflow consisting of five distinct phases: engine foundation, system building, integration, and a final critique loop. The critic agents, acting as art directors, evaluate screenshots of the running application against specific criteria such as material quality, lighting, and motion. If the output fails to meet a target score, the feedback is fed back into the primary agent to trigger a new iteration.

Autonomous Tooling and Harness Creation

During the 19-hour generation process for the Formula 1 game, the model autonomously constructed a custom testing harness. This included the creation of 136 single-purpose tools, such as a road material A/B testing tool and a scripted input driver for car movement. The model also developed three permanent tools to facilitate the critique loop: a screenshot utility, a state extractor for shadow and lighting data, and a regression-testing tool that compares screenshots between iterations to ensure consistency.

Performance and Resource Usage

The experiment consumed 1.7 billion tokens across 137 agents, resulting in a total cost of $1,200. The model reached a peak quality score of 67.3 out of 100 before plateauing. While the resulting game, Apex GP, features functional UI, mini-maps, and basic physics, it suffers from collision detection bugs and lacks a formal end-of-race state. The generated code is entirely self-contained, requiring no external assets, and demonstrates the model's ability to manage complex 3D scene generation and integration tasks without human intervention in the codebase.