AI Coding Agents Boost Code Output, Not Software Production

AI coding agents are highly effective at generating large volumes of functional code, but their overall impact on software production efficiency is limited by the human code review process. A recent study conducted by researchers at Harvard University, Fiona Chen and James Stratton, analyzed aggregated analytics data from Jellyfish, a platform that tracks engineering team output. This data included 300 million "work events," such as code commits and pull requests, along with issue management software data. The dataset spanned over 700,000 employees across more than 700 software development firms, covering the period from 2021 through March 2026. The study's findings indicate that while AI tools accelerate the initial coding phase, this efficiency gain is absorbed by downstream constraints. Specifically, the research observed that the code review process significantly lengthens, pull requests are more prone to requiring revisions, and reviewers tend to leave a greater number of comments. Consequently, the study found "little evidence that firms increase software output or reduce employment" as a direct result of adopting these AI coding tools. The researchers concluded that the human code review stage acts as a substantial bottleneck, preventing the full realization of efficiency gains promised by AI-assisted code generation. This suggests that current AI coding assistants, while proficient in producing code, do not fundamentally alter the overall software development lifecycle in a way that dramatically increases the volume of delivered software or reduces the need for human oversight and quality assurance. The study's methodology involved examining granular output data to understand the practical implications of AI adoption in software engineering workflows. The aggregated data provided a broad view of how AI coding tools are integrated into real-world development environments and their measurable effects on team productivity and output. The implications of these findings are significant for companies investing in AI coding assistants, highlighting the need to address the entire development pipeline, including review processes, to fully leverage the potential of these technologies. The research underscores that the efficiency of AI-generated code is heavily dependent on the capacity and structure of the subsequent stages of software development, particularly the critical human element of code validation and integration. The study's timeframe, from 2021 to March 2026, provides a contemporary perspective on the evolving landscape of AI in software development, capturing trends and impacts over several years. The use of data from over 700 firms and 700,000 employees offers a robust sample size for drawing conclusions about the widespread adoption and effects of AI coding agents in the industry. The researchers' focus on "work events" and issue management data allowed for a detailed analysis of the development workflow, from initial code creation to final deployment considerations. The study's emphasis on the "bottleneck" of code review points to a key area for future improvement and optimization in the integration of AI into software engineering practices.