Your review bottleneck is an abstraction problem
Agents generate code faster than humans can review it, and the popular fix is a second model. But review is slow because of what review actually is, and the only real lever is making the change small enough to understand.