AI coding agents generate more code, but not more software
Points and comments are a snapshot, not live.
AI coding agents boost code output 30% but human review bottlenecks limit software gains.
A study of 700+ firms by Harvard researchers Fiona Chen and James Stratton finds AI coding agents increase lines of code by 30% and commits by 20%, but resolution of software features (Issues/Epics) shows no statistically significant change. Code review time balloons 49% on average, pull requests requiring changes nearly double, and comments per PR rise 35%. Firms added 14% more workers to code review but saw no significant employment changes. The researchers conclude coding efficiency gains are "absorbed by downstream constraints" in the review process.
AI agents generated only 23.3% of review comments and 10.8% of pull requests by March 2026, leaving most review work to humans. The authors note the tools are still new and trade-offs could improve with experience.
What commenters are saying
Many commenters strongly agree with the study's findings from direct experience, describing "soooo much slop code" that must be sifted through. A senior engineer notes that code reviews remain the "most tedious part of my job."
Some point out benefits: one says AI catches bugs that previously went unnoticed, though sorting through many false positives adds work. Others argue code review's value is shared understanding, not just quality gates. A few suggest the path forward is startups adopting no-human-review policies from day one, which would force the industry to adapt if successful. A commenter observes that automated test suites should catch performance regressions, making line-by-line review less critical in those areas.