fion.ac (Harvard Economics, Fiona Chen & James Stratton) web signal

Harvard Job Market Paper: AI Agents Raise Individual Coding Output 30% But Firm Software Output Doesn't Move — Code Review Is the Bottleneck

By Alexis Dufresne Published October 21, 2026 at 12:38 UTC Updated October 21, 2026 at 12:41 UTC

Summary

Causal study of 718 firms and 725,938 workers (300M work events, Jan 2021–Mar 2026) finds AI coding agents boost individual coding activity 30% in lines of code and 23% in pull requests, yet firm-level software output shows a statistically zero effect — ruling out gains larger than 12% of baseline, well below the individual productivity spike. The culprit is code review: review time rises 49%, the share of PRs requiring revisions nearly doubles, comments per PR rise 35%, and firms are forced to redeploy engineers toward review rather than shipping; employment holds flat at zero, equally dismantling the layoff narrative.

Shared on Bluesky by 5 AI experts

Dan Greene @dmgreene.bsky.social: Fiona Chen's job market paper appears to show that ai coding assistant adoption increases coding productivity but without full pass-through … →
Dare Obasanjo @carnage4life.bsky.social: The solution the industry seems to be landing on is throwing more AI at the problem. AI writes the code, AI reviews the code and no one know… →
samuel mehr @mehr.nz: it'll be really interesting to see if the effects hold up to scrutiny in peer review (this is the first author [Chen's] job market paper, un… →
Matthew Guzdial @matthewguz.bsky.social: I think the crash won’t directly impact prices much, DeepSeek is likely still around even post crash and that’s about as good as you can get… →
Joshua Grochow @joshuagrochow.bsky.social amplified

@shimminykricket.blacksky.app

Harvard just dropped a study AI productivity in software engineering and it found the average increase in productivity from using Claude code is.....zero and for exactly the reasons you'd think fion.ac/jellyfish.pdf
View on Bluesky →

Originally reported by fion.ac (Harvard Economics, Fiona Chen & James Stratton)

Read the original article →

Original headline: Harvard Job Market Paper: AI Agents Raise Individual Coding Output 30% But Firm Software Output Doesn't Move — Code Review Is the Bottleneck