Software engineering has long run on strong opinions and weak evidence. AI coding agents may finally make it possible to study how software is built with real experimental rigor.