From Ars Technica · · 2m
AI coding agents write more code, not better software
A study of 700,000 employees finds AI coding tools boost output but get bogged down in human code review.
In brief
A study of 700,000 employees finds AI coding tools boost output but get bogged down in human code review. AI coding agents increase code volume by 30 percent but do not improve software output because human code review becomes a major bottleneck. Originally reported by Ars Technica.
AI code generation works, but trust is the problem
Modern AI coding assistants generate huge amounts of functional code efficiently. But programmers know they cannot trust the accuracy, so substantial effort reviewing AI output is required.
Study finds code review is the bottleneck
Harvard researchers Fiona Chen and James Stratton studied 300 million work events across over 700,000 employees at more than 700 software firms from 2021 through March of 2026.
AI agents boost raw code output sharply
Introduction of AI coding agents led to a 30 percent increase in total lines of code, a 20 percent rise in commits, and a 23 percent increase in pull requests on average.
But software output does not improve
The resolution rate for issues and epics tracked by tools like Jira did not change in a statistically significant way after AI tools were introduced.
Code review time balloons by nearly half
The average review process time between pull request submission and merge balloons 49 percent on average after AI agents are introduced.
Reviewers request changes far more often
The share of pull requests with changes requested nearly doubled, and the number of comments per pull request increased by 35 percent following AI agent introduction.
More reviewers needed, but no job losses yet
A 14 percent increase in workers performing code reviews followed AI agent introduction. Researchers found no significant employment changes attributable to AI after cross-referencing with LinkedIn data.
AI review tools have barely helped so far
Although 80 percent of measured firms used some form of AI code review by March 2026, AI agents were only responsible for 23.3 percent of review comments and 10.8 percent of pull requests.
Teams still learning how to use AI coding agents
Ninety-five percent of firms have implemented AI coding agents, but many are still learning when and how to deploy them best. Trade-offs between coding and review time could improve with experience.
What it means
AI coding agents increase code volume by 30 percent but do not improve software output because human code review becomes a major bottleneck.
