This matches my experience. For a lot of agent-written code, reviewing the final diff feels too low-level and too late. I'd rather review the plan&assumptions.

This matches my experience. For a lot of agent-written code, reviewing the final diff feels too low-level and too late. I'd rather review the plan&assumptions. If those are right, AI will probably get the implementation right as well.