Harness-of-Harness lifts coding agents 52%

The important idea isn’t simply running an agent longer; it’s creating an evidence-carrying control loop that preserves verified progress between sessions. The results are promising, but the headline gains remain research claims rather than proof of production-ready autonomy.

• –Reports a 52.25% average relative gain and up to 82.86% after three iterations across three benchmarks
• –Separates planning, development, and independent testing to reduce regressions and self-evaluation bias
• –Demonstrates long-horizon capability by building a playable FPS across more than 70 autonomous loops
• –Keeps versioned artifacts, issue histories, and evidence packets so each iteration can build on validated work
• –HoH-lite and reproducibility materials are planned, which should make the claims easier to independently verify