AI coding agents increased coding activity; study reports more review work

Chen and Stratton studied how AI coding assistants and agents affected software-engineering work, using activity data from an engineering analytics platform.[1][2] The dataset contained 300 million work events across 718 firms, including GitHub activity, Jira issues and Google Calendar events.[2] It covered January 2021 through March 2026.[2] The researchers used a staggered difference-in-differences design to compare outcomes around firms’ adoption of AI coding tools.[2] The paper’s current version is dated August 4, 2026.[2]

Following firms’ adoption of AI coding agents, lines of code increased 30%, commits 20% and pull requests 23% on average.[2] For comparison, AI coding assistants were associated with increases of 12% in lines of code, 9% in commits and 5% in pull requests; only the commit increase was statistically significant.[2] The paper describes agents as tools that can semi-autonomously complete tasks of significant length and complexity.[2] The researchers’ estimates rule out an increase in software output greater than 12% of the baseline mean for agents.[2]

The share of pull requests with changes requested nearly doubled.[2] Comments per pull request increased 35%, and the share of workers performing code reviews increased 14%.[2] The study reported these increases in review activity; the researchers said human reviewers remained central to the process.[2] For teams assessing the practical effect of agents, the findings suggest that output counts alone may not capture the additional work involved in checking and revising code.[2]

The researchers found no statistically significant effect on employment after firms adopted AI coding tools.[2] For AI agents, they estimated a precise zero effect on overall employment and a less precise zero effect on engineering employment.[2] By March 2026, 95% of firms in the study had implemented AI coding agents.[1] At that point, 80% used some form of AI code review, while agents produced 23.3% of review comments and 10.8% of pull requests.[1] The paper says its conclusions are the authors’ own and do not represent Jellyfish’s views.[2]