Original title: Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
Haoyang Yan, Min-le Su, Hangfan Zhang, Zhanhao Li +5 · September 2026
Turning broad ideas into finished software without human help is hard for AI coding agents. To solve this, researchers built Harness-of-Harness, or HoH. This system runs existing coding tools through repeated planning, coding, and testing loops so they can keep improving their work over time. It breaks tasks into small, checkable steps, balances bug fixes with adding new features, and tests code independently. It also saves past versions, shares tools as needed, and reuses old code instead of starting over. Tested across three benchmarks, HoH beat standalone systems, achieving an average relative gain of 52.25 percent after three iterations. In a multi-day test with over 70 iterations, it built a complete, playable first-person-shooter game with a coherent story, polished visuals, and sound.
Imagine you are learning to paint a picture. At first, you sketch a rough outline, then add colors, but some parts don't look right. You step back, adjust the colors or shapes, and try again, each time making small improvements until the whole painting comes together. In this paper, the AI coding agent does something similar: it starts with basic code, then through repeated cycles of planning, coding, and testing, it refines the software incrementally until it becomes a complete, functional game.
This matters because it shows how AI can autonomously develop complex software without human help, potentially speeding up creation of games and applications. Developers could use such frameworks to automate parts of software engineering, reducing time and cost.
The abstract does not mention specific limitations, but it focuses on demonstrating improvement in coding benchmarks and game development. It does not address scalability to other software domains or potential failures in very large or safety-critical systems.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement — Haoyang Yan, Min-le Su, Hangfan Zhang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.