Harvard Study Finds AI Agents Fail to Boost Software Output

A recent study by Harvard University researchers Fiona Chen and James Stratton reveals that automated AI coding agents increase code volume without delivering a corresponding rise in completed software features. Examining aggregated engineering telemetry from Jellyfish—spanning 300 million work events across more than 700,000 workers at over 700 software development companies between 2021 and March 2026—the authors analyzed the real-world operational effects of incorporating AI coding tools and autonomous agents into engineering workflows.

The research shows that deploying autonomous AI coding agents causes total generated lines of code to surge by 30 percent, while overall commits rise by 20 percent and pull requests grow by 23 percent. Despite this swell in raw output, the resolution rates for actual software features, such as Issues and Epics tracked in management tools like Jira, showed no statistically significant improvement. Furthermore, cross-referenced data from LinkedIn showed that adopting these AI tools resulted in no noticeable changes in overall company employment levels.

This discrepancy stems from a dramatic slowdown during the human review stage. The average turnaround time between submitting a pull request and merging it into a codebase ballooned by 49 percent following agent adoption. Additionally, the proportion of pull requests requiring revisions nearly doubled, and reviewer comments per pull request jumped by 35 percent. To manage the heavier evaluation workload, organizations saw a 14 percent increase in the share of staff assigned to review code. Although 80 percent of companies implemented automated code review tools by March 2026, AI handled only 23.3 percent of review comments and 10.8 percent of pull requests, leaving human developers burdened with the vast majority of verification work.

For software practitioners, these findings indicate that accelerating initial code generation merely shifts engineering constraints downstream. Rather than speeding up product delivery, agentic assistants demand greater time and effort during verification to filter out low-quality contributions. Until automated review systems mature, engineering teams must balance raw generation speed against the added evaluation burden imposed on human reviewers.