Study: AI Coding Agents Boost Code Output but Not Software Delivery

A new study from Harvard University researchers Fiona Chen and James Stratton finds that AI coding agents produce substantially more code, but that gain does not translate into more finished software or fewer jobs. The research, based on aggregated analytics from Jellyfish, covers 300 million work events such as commits and pull requests across more than 700,000 employees at over 700 software development firms from 2021 through March 2026.

Code Output Rises Sharply

The study measured when each company introduced AI coding assistants, which help auto-complete primarily human-authored code, and AI coding agents, which write and submit code autonomously based on prompts. Using a difference of differences regression on key variables before and after adoption, the researchers found that introducing AI coding agents leads to a 30 percent increase in total lines of code generated, a 20 percent rise in the number of commits, and a 23 percent increase in pull requests on average.

Reviews Become the Bottleneck

That extra code does not become extra software. The resolution rate for Issues and Epics, meaning wholesale software features tracked in tools like Jira, showed no statistically significant change after AI tools were introduced, and the researchers found no shift in the size or complexity of those tracked items. The reason lies in the human code review process. The average time between a pull request being submitted and merged into the codebase grew 49 percent after AI agents were introduced. The share of pull requests with changes requested nearly doubled, and the number of comments per pull request increased by 35 percent. In response, the share of workers performing code reviews rose 14 percent.

Little Impact on Output or Jobs

The researchers also found no significant employment changes attributable to AI, based on total active workers in the Jellyfish data cross-referenced with LinkedIn records at those firms. AI has so far played only a marginal role in review work: although 80 percent of measured firms used some form of AI code review by March 2026, AI agents accounted for just 23.3 percent of all review comments and 10.8 percent of all pull requests, leaving humans responsible for the vast majority of that work. With 95 percent of firms in the study having implemented AI coding agents by the data cutoff, the authors suggest that coding time versus review time trade-offs could improve as engineering teams gain experience with when and how to deploy the tools. For now, the study frames AI coding as a double edged sword, with increases in coding speed counteracted by similar increases in human code review time and effort.

Sources: Arstechnica