AI coding agents add 30% more code, not software output

AI coding agents increased code volume by 30% without measurable software delivery gains, according to a study of more than 700 firms.

Harvard University researchers Fiona Chen and James Stratton reported that review periods lengthened after adoption, while completed software work showed no statistically significant increase. Their findings distinguish faster code generation from the broader task of delivering finished software.

Review workloads rise after adoption

The researchers estimated that agent adoption increased commits, or recorded code changes, by 20% and pull requests by 23%, alongside the increase in code volume. Pull requests submit proposed changes for examination before they enter the shared codebase.

That examination took longer: the average interval between submission and integration increased by 49% following adoption, the researchers reported. Requests requiring revisions became nearly twice as common, while comments on each request increased by 35%.

Chen and Stratton attributed the gap between code generation and finished work to constraints later in production, particularly review. They also estimated a 14% increase in the proportion of employees undertaking reviews, suggesting the additional output brought a redistribution of work rather than simply removing tasks.

Jellyfish records track completed work

The study drew on engineering analytics provider Jellyfish, covering 300 million recorded activities across more than 700,000 employees. Its observation period ran from 2021 through March 2026, encompassing code activity and task-management records at the participating businesses.

The researchers distinguished assistants, which support human-written code, from agents that can produce code and submit changes in response to instructions. They combined observed AI use with GitHub activity to identify adoption timing, then compared changes across organizations before and after deployment using a difference-in-differences analysis.

To assess delivery rather than typing volume, they examined completion of tracked tasks and larger feature groupings in systems including Jira. Neither measure showed a statistically significant improvement after adoption, and the researchers found no corresponding change in task size or complexity that explained the result.

The employment findings were similarly limited: Chen and Stratton did not establish a statistically significant staffing change attributable to AI. They examined active-worker records from Jellyfish alongside LinkedIn information; that finding does not establish that individual jobs or responsibilities remained unchanged.

Automated review remains a partial answer

By March 2026, 80% of the measured companies had adopted AI-assisted review in some form, according to the researchers. Yet automated agents supplied only 23.3% of review comments, leaving people responsible for most of that recorded activity.

The distinction matters for software businesses evaluating these tools: code volume, review effort and completed features measure different stages of production. The findings do not establish that generating additional code is worthless, but they provide no statistically significant evidence of higher delivery rates across the observed firms.

If review constraints persist, an adopting company could produce more proposed changes without completing more features, limiting delivery gains across the sector. If teams improve checking and revision workflows, more of the additional code could reach finished products; these are conditional possibilities, not outcomes established by the study.

The March cutoff leaves later tools and working practices outside the evidence. Whether subsequent improvements shorten reviews while increasing completed work remains unresolved, and the study does not establish an economy-wide productivity effect.