AI coding agents generate more code, but not more software

A recent study by Harvard researchers Fiona Chen and James Stratton, utilizing data from over 700,000 employees across 700 firms, reveals that AI coding assistants have failed to increase overall software output. While these tools significantly accelerate the generation of functional code, the study identifies human code review as a critical bottleneck. The efficiency gains achieved during the initial coding phase are largely offset by the increased time and effort required for downstream review processes. Data from Jellyfish, covering 300 million work events between 2021 and 2026, indicates that pull requests generated by AI are more likely to require revisions and receive more reviewer comments. Consequently, the research finds little evidence that firms are increasing their total software production or reducing employment, as the production process remains constrained by the capacity of human developers to verify and integrate AI-generated output.