Mert Demirer at MIT Sloan, with Leon Musolff and Liyuan Yang, studied more than 100,000 developers on GitHub adopting three kinds of AI coding tools. The numbers step down as you move toward the user. Autocomplete raised coding activity about 40%. Adding synchronous agents took it to 140%. All three tool types together reached 180% more coding output. Then: 50% more projects, 30% more releases, and no detectable increase in downloads or reviews of the apps that shipped.
The researchers' explanation is the one I would have guessed. The human processes downstream of writing code did not change. Review, testing, polishing, and release took the time they always had, so the output of the fast step piled up in front of the slow ones. The gain was real at the keyboard and mostly gone by the time anything reached a person.
This is the third large study this year to find the same shape, and the shape is the finding. The tools moved the bottleneck from writing code to everything after it. The organizations that will see the gain are the ones that move with it: more review capacity, faster test infrastructure, release engineering treated as a first-class job rather than the thing that happens at the end.
An engineering leader should stop asking whether the tools make developers faster. They do. The question is what the organization is going to do about the step that is now the slowest, and whether anyone owns it.
Where does work pile up on your team now, and who owns that step?
Photo source: https://photos.robertstowe.com/san-diego