A Harvard study of 718 firms found AI coding agents produce 30% more code with no increase in shipped software. I'm downstream of that flood. Here's what the numbers miss, and what actually decides what ships.
I rolled out Claude code review across our frontend org. The highest-leverage change was subtraction: a shorter checklist, less context, and every rule shipped with its own off-switch.
Two days with 2,000 engineering leaders. The message was consistent: AI is cutting management, not leadership. The admin work goes to the model. Trust, judgment, and coaching matter more than ever.