Addy Osmani, Morgan Westlee Lunt, and Qing Wang described a reconception of review as a continuous process rather than an end-of-process gate
Following·Day 9·5 sources
Harvard Study: AI Boosts Code 30% but Shipped Software Unchanged
The gist
The study shows that raw code volume gains from AI agents do not translate to more shipped software at the firm level. Review capacity, not generation capacity, is the binding constraint.
The full picture
A Harvard economics study of 718 firms and roughly 700,000 workers, conducted by Fiona Chen and James Stratton, found that AI coding agents increased lines of code written by 30% but produced no meaningful increase in completed tasks, shipped software, or employment levels. Code review duration grew 49%, the share of pull requests requiring changes nearly doubled, and reviewer comments per pull request rose 35%. The bottleneck in AI-assisted software development shifted from code generation to code review.
How it developed
ComplexDiscovery reported the study's scope of roughly 700 firms and 700,000 workers, and the AI authorship metadata governance angle
DavidLinthicum reported the Harvard study findings including unchanged firm-level output and employment, and the PR rework metrics
Sources
Simon Willison
complexdiscovery.com
The daily email
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free