Celadon Global / Unsplash AI coding tools tripled output on GitHub. Releases rose 30%. A Wharton study says reviews are the bottleneck.
Wharton and MIT researchers tracked 100,000 GitHub developers from 2022 to 2026 and found each generation of AI coding tools boosted coding activity sharply — while the gains in finished, released software lagged far behind.
The numbers on AI-assisted coding keep getting bigger, and a new economics study suggests the impressive part is also the easy part. Writing up research by Wharton professor Leon Musolff and MIT researchers Mert Demirer and Liyuan Yang, Knowridge reports that the team tracked more than 100,000 GitHub developers from 2022 to 2026, comparing productivity before and after they adopted three successive generations of AI coding tools — and matched public GitHub records with Microsoft data on when developers actually started using the tools.
The headline figures are striking. Autocomplete systems that suggest the next line of code increased coding activity by 40%. Adding “sync agents” that edit code alongside developers in real time lifted the cumulative increase to 140%, and “async agents” that work autonomously from a prompt pushed it to 180%. Every generation of tool compounded the gains at the level of individual coding work.
But the same study found those gains attenuating sharply downstream. The 180% cumulative jump in coding activity translated into only a 50% increase in software projects and a 30% increase in software releases. “In software, the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it,” the authors write. The bottleneck moved from the keyboard to the human processes around it: code review, integration, coordination, distribution.
The pattern held outside GitHub as well. The researchers examined the four biggest software marketplaces — the Apple App Store, Google Play, the Chrome Web Store, and SourceForge — and found a surge of new applications since mid-2025 with no corresponding increase in overall usage. On the App Store, monthly new releases rose from roughly 30,000 before AI coding agents arrived in early 2025 to about 100,000 per month by April 2026, while total usage stayed flat or declined. Musolff suggests discovery is part of the problem: “It could simply be that it’s much harder to discover new applications when there’s such a flood of them.”
Musolff’s take is blunt about the present and cautiously optimistic about the trajectory: “If the world froze at today’s level of AI capabilities, these results would be a bit of a cold shower.” Still, he notes that each generation of tools has attacked a later stage of the development process, so the gap between coding productivity and shipped software could narrow as the tools improve — and few other technologies deliver a 30% increase in releases at all. One caveat he flags is pointed for the current wave of AI-for-code-review products: “If the same AI that wrote the code also reviews it, that doesn’t really solve the problem. The review just isn’t of the same quality.”
A few notes on provenance. This is a working paper, not a peer-reviewed publication, and the figures above come from the version summarized in the current coverage; the author has since posted a revised version with expanded data. Also worth noting: the researchers worked with Microsoft telemetry to identify tool adoption timing, so the Microsoft ecosystem is overrepresented in the measurement. Even with those caveats, the direction is hard to dismiss — the most rigorous attempt yet to measure AI coding tools against what actually ships, and the answer is that shipping was never mostly about writing.