Model-Agent Flywheel

The flywheel spins fastest where one player holds both ends (Anthropic + Claude Code): the model knows how Harness will call it, Harness knows the model’s boundaries. Experiment: without model changes, only changing Harness, accuracy increased from 52.8% to 66.5%—Harness leverage is high precisely because the model hasn’t reached this point yet. The flywheel is the deepest moat of the era. For model builders: spin real-world feedback into training. For application developers: Harness is a sharp lever in the short term, but the durable moat is beyond technology: proprietary data, channels, user trust, network effects. The three questions “what it sees, what it can do, how to verify” will not become obsolete.

Related: [Coevolution of Model and Agent], [Harness Engineering], [Agent Formula]