The flywheel is a fascinating framework for driving incredible breakthroughs. We see it in DL & RL (dataset → model → evaluation → update), in Lean Startup (build → measure → learn), and even in Popper's philosophy of science (hypothesis → prediction → explanation → new hypothesis). For the flywheel of autonomous AI, the bottleneck inevitably lies in evaluation. As we build AI agent systems, we are moving toward 'expert-in-the-flywheel'—where domain expertise becomes more vital than ever.