AI coding agents generate more code, but not more software

A study of hundreds of software firms found that AI coding agents generate significantly more lines of code and commits, but this increased output does not translate into more completed software features or reduced employment. The efficiency gains from AI writing code are offset by a substantial bottleneck in the human code review process, which takes 49 percent longer on average after AI agents are introduced, with reviewers requesting changes on nearly twice as many pull requests. While AI tools are now widely adopted across firms, humans still perform the vast majority of code review work, suggesting that the time saved in the coding phase is being consumed by the review phase instead.