Shanghai AI Lab's Harness-of-Harness runs 70+ iterations to build an FPS game autonomously
aigclink · x · 2026-09-05
Shanghai AI Lab released Harness-of-Harness (HoH), a meta-framework layered on existing coding-agent harnesses that organizes execution into iterative planning-coding-testing loops. Each round scopes a small verifiable increment and passes code plus test evidence to the next, preventing memory loss and premature completion claims.
- Three roles (Project Planner, Developer, QA Tester) are three constrained calls to the same coding agent
- Across GameCraft-Bench, FrontierSWE, and ProgramBench with three harness-model pairs, HoH gains 52.25% average relative improvement (max 82.86%) after three iterations
- In a multi-day deployment, HoH autonomously built Fusepoint, a playable FPS with coherent story, mechanics, visuals, and audio, in 70+ iterations from a single PRD
Related event: Shanghai AI Lab's Harness-of-Harness builds a playable FPS autonomously(2 posts)→
More from coding & agent
- GPT-6 Astra builds a procedural three.js game under 2MB, then remakes it in Blender — gabrielchua · 2026-09-05
- OpenAI brings ChatGPT and Codex into UPenn drug discovery workflows — AllThingsApx · 2026-09-05
- Codex /fast mode runs at 2x speed and stays within subscription, unlike Claude's paid extra usage — rudrank · 2026-09-05
- ChatGPT Workspace 5-hour limit mid-agent-run wipes uncommitted code, dev demands fix — cneuralnetwork · 2026-09-05
- Astra costs ~$1.49/M tokens for this dev's coding — but is it good enough to be the default? — notchillatall · 2026-09-05
- HeroUI launches Agents: embed a context-aware agent in your product — ycombinator · 2026-09-05