A coding agent just spent several days building a first-person shooter from a product brief and an empty workspace. After more than 70 planning, coding and testing loops, it had produced a playable game with a storyline, mechanics, visuals and audio. No human intervened in the code. The framework, Harness-of-Harness, wraps existing coding agents in a recurring loop that tests small increments and preserves a versioned project history. Across three benchmarks and three model-harness pairs, three iterations improved performance by 52% on average over the standalone harnesses, with a maximum gain of 83%. While this doesn't prove agent loops will keep improving a product indefinitely, it does highlight some important gaps in review changes when agents work across days. A finished product can contain dozens of autonomous decisions nobody inspected individually. When it comes to eventual debugging, every decision, test and change will need a traceable record because most agents can't backtest their work and figure out which edits broke what.(Sahara AI 🔆)