Harness-of-Harness framework boosts autonomous coding agents by 52% over multi-day runs
rohanpaul_ai · x · 2026-09-06
An arXiv paper introduces Harness-of-Harness (HoH), a framework that lets LLM coding agents autonomously develop and continually improve software over multi-day runs without human intervention.
- HoH wraps existing coding-agent harnesses and organizes execution into iterative planning-coding-testing loops
- Key design: balancing repair with capability growth, small verifiable increments, separating implementation-time testing from independent evaluation, versioned project histories, and reuse over recreation
- Across GameCraft-Bench, FrontierSWE and ProgramBench, three harness-model pairs (Codex+GPT-5.5, OpenCode+DeepSeek-V4-Pro, Pi+MiniMax-M3) consistently beat standalone harnesses, averaging a 52.25% relative gain and up to 82.86% after three iterations
- In a 70+-iteration multi-day deployment, HoH autonomously developed a complete first-person game
The core argument: long-horizon coding needs persistent project state, independent testing, and replanning from real failures—not just more time.
More from coding & agent
- GLM Coding Plan ups Flash quotas: unlimited in ZCode, 2x elsewhere — pcuenq · 2026-09-06
- Do production AI agents actually need an 'Agent SRE'? A developer asks to be proven wrong — Fantastic-Sleep-3352 · 2026-09-06
- Grok auto-suggests connectors and hot-swaps them into agent context without restart — Baconbrix · 2026-09-06
- Dev uses Astra on ultra to blueprint systems and build a game's first vertical slice — Dimillian · 2026-09-06
- OpenAI researcher: old Skills now hurt GPT-6 Astra — audit and clean up your AGENTS.md — udmrzn · 2026-09-06
- Dev builds and open-sources a Codex Micro-style hardware terminal for coding agents — VoidStateKate · 2026-09-06