The question

Ars Technica reported on Friday that Harvard University researchers Fiona Chen and James Stratton looked at actual development inside firms, not lab exercises. Coding agents add code. Software delivery, measured by issues, does not rise with it. They place the gap in longer human review.

[1]
An open laptop with a dark screen beside a stack of folders, one with an oxide-red tab, and no readable letters on the keys
The gouache shows a dark-screen laptop and a stack of folders. It stands for the longer review that follows the extra code, and it is not a news photograph., AI-generated illustration, not a news photograph

Where the data come from

The researchers used aggregated analytics from Jellyfish: 300 million work events, such as commits and pull requests, plus issue-tracker data, across more than 700,000 employees at over 700 software firms, from 2021 through March 2026.

They used directly measured AI use and an analysis of GitHub activity to judge when each firm started using coding assistants or coding agents. Assistants mainly complete code a person is writing. Agents mainly write and submit code from prompts. The researchers then ran what Ars calls a difference of differences regression, comparing firms before and after they introduced the tools.

[1]

More code, not more finished work

After a firm introduced coding agents, lines of code rose 30 percent on average, commits rose 20 percent, and pull requests rose 23 percent. The resolution rate for Issues and Epics tracked in tools such as Jira did not change in a statistically significant way. The researchers also found no compositional shift in the size or complexity of those issues.

Review got longer. The average time from submitting a pull request to merging it rose 49 percent. The share of pull requests with changes requested nearly doubled, and comments per pull request rose 35 percent. The share of workers doing code review rose 14 percent. After checking active workers in Jellyfish against LinkedIn data at those firms, the researchers write that they cannot attribute significant employment changes to AI.

By March 2026, 80 percent of the measured firms used some form of AI code review. Even so, AI agents wrote only 23.3 percent of review comments and were involved in only 10.8 percent of pull requests. People still did most of the review.

[1]

What the report cannot show

Ars does not give a paper title or a preprint link, and this article does not invent one. The measurement window ends in March 2026. Ars also writes that the tools have been updated since then, that 95 percent of firms in the study have implemented coding agents by this point, and that many teams are still learning when to use them. The article does not date that 95 percent figure on its own. Longer review and the issue-resolution result are averages on this sample, not a finding about every firm.

[1]