A recent study from Harvard University has revealed that while AI coding agents can churn out vast amounts of code, the quality bottleneck remains firmly in human hands. The research, which analyzed data from over 700,000 employees across more than 700 firms, found that the introduction of AI coding tools led to a 30% increase in lines of code, yet there was no significant increase in software output or reduction in employment.

The key finding is that while AI can generate large volumes of code, the rigorous code review process required to ensure quality is a bottleneck. This means that even with AI’s help, the overall efficiency gains are minimal. The researchers noted that pull requests often required revisions and reviewers left more comments, indicating that the review process is significantly more time-consuming.

The study’s authors, Fiona Chen and James Stratton, used data from engineering teams and GitHub activities to pinpoint the exact impact of these tools. They found that while the number of commits and pull requests increased, the resolution rate for software issues, such as Jira-tracked issues, did not change. This suggests that AI tools are valuable for code generation but not in reducing the amount of human oversight needed.

Despite the efficiency gains during the coding phase, the study highlights that the downstream constraints of code review negate these benefits. This means that while AI can help in writing more code, it cannot replace the crucial role of human coders in ensuring quality and functionality.