Study Finds AI Coding Agents Increase Code Generation but Not Software Output

<p>AI coding assistants and agents are efficient at generating large amounts of functional code, according to industry observations. However, developers express concerns about the accuracy of this code, necessitating significant effort to review AI-generated output.</p><p>A recent study conducted by researchers from Harvard University analyzed coding practices across hundreds of firms. The study found that human code review creates a significant bottleneck in the efficiency of AI coding tools, leading to minimal evidence that firms increase software output or reduce employment by using them. The efficiency gained during coding is countered by constraints in the production process; as the code review process lengthens, pull requests are more likely to require revisions, and reviewers tend to leave more comments.</p><h2>Cut once, measure twice</h2><p>The researchers, Fiona Chen and James Stratton, utilized aggregated analytics data from Jellyfish, which tracks the detailed output of engineering teams. This data includes 300 million individual work events, such as commits and pull requests, along with issue management software data from over 700,000 employees at more than 700 software development firms from 2021 through March 2026.</p>