From zero to a playable FPS

Fusepoint is a single-player narrative first-person shooter developed by HoH from an empty workspace containing only a user-provided product requirements document (PRD), with no human in the loop.

Abstract

This paper studies autonomous software development by LLM-based coding agents: agents autonomously turn high-level requirements into complete, functional, and usable software without further human intervention.

Harness of Harness (HoH) builds on existing coding-agent harnesses and organizes their executions into iterative planning–coding–testing loops. It balances repair with capability growth, scopes development into small and verifiable increments, separates implementation-time testing from independent evaluation, and constrains verifiable outputs rather than prescribing agent workflows.

Across GameCraft-Bench, FrontierSWE, and ProgramBench, and three harness–model pairs, HoH achieves an average relative gain of 52.25% and a maximum gain of 82.86% after three iterations. In a multi-day deployment with more than 70 iterations, HoH autonomously develops Fusepoint, a human-playable first-person shooter with a coherent storyline, fully implemented core mechanics, polished visuals, and integrated audio.

Method

Harness-of-Harness framework with project planning, development, and evidence-grounded QA testing

Results

We evaluate Vanilla and HoH@1–3 on three benchmarks—GameCraft-Bench, FrontierSWE, and ProgramBench—using Codex with GPT-5.5 (high), OpenCode with DeepSeek-V4-Pro, and Pi with MiniMax-M3.

Harness + Model	Setting	GameCraft-Bench	FrontierSWE	ProgramBench
Harness + Model	Setting	Overall	Dominance^‡	Pass Rate^†
Codex GPT-5.5 (high)	Vanilla	49.58	44%	60.41
	HoH@1	59.71	58%	65.42
	HoH@2	64.84	60%	65.79
	HoH@3	71.52(+21.93)	71%(+27)	66.50(+6.09)
OpenCode DeepSeek-V4-Pro	Vanilla	26.90	25%	45.27
	HoH@1	28.61	28%	55.33
	HoH@2	40.32	42%	55.66
	HoH@3	48.98(+22.08)	44%(+19)	57.56(+12.29)
Pi MiniMax-M3	Vanilla	42.16	35%	35.83
	HoH@1	49.06	62%	48.68
	HoH@2	55.04	66%	53.57
	HoH@3	58.78(+16.62)	64%(+29)	52.68(+16.85)

Main results on GameCraft-Bench, FrontierSWE, and ProgramBench. Each harness–model pair is evaluated under Vanilla and HoH@1–3. Bold values mark the best setting within each harness and metric. Green values shown only for HoH@3 give absolute gains over Vanilla; Dominance gains are percentage points and higher is better for all metrics. ^‡ Dominance is an official FrontierSWE metric; see the official documentation for details. ^† Pass Rate denotes ProgramBench's average test pass rate.

Conclusion

Harness of Harness treats greenfield software development as a harness-level, long-horizon research problem. Across the controlled benchmarks, HoH improves final artifact quality for all three harness–model configurations and continues to benefit from additional iterations. In the multi-day Fusepoint case, HoH transforms a PRD and an empty workspace into a complete, human-playable FPS over more than 70 loops, while versioned artifacts, issue histories, and evidence packets preserve the development trajectory. Future work will extend HoH to broader real-world scenarios, including different types of games and other software systems.

Resources

We will publicly release HoH-lite, a lightweight implementation of HoH's core workflow, together with reproducibility materials.

Citation

The paper is available on arXiv:2609.01481.

@article{yan2026harness,
    title={Harness of Harness: Multi-Day Autonomous Software Development with Continual Improvement},
    author={Yan, Haoyang and Su, Min-Le and Zhang, Hangfan and Li, Zhanhao and Zhang, Chen and Zhang, Shao and Chen, Yang and Bai, Lei and Hu, Shuyue},
    journal={arXiv preprint},
    eprint={2609.01481},
    archivePrefix={arXiv},
    year={2026}
}