An analysis of more than 700 firms finds longer code reviews, more revisions and no significant staffing changes after adoption

AI coding agents increase the volume of code produced, but a study of more than 700 software firms has found no statistically significant rise in completed software work. Instead, faster coding appears to shift effort downstream: human reviewers spend more time checking and revising AI-generated contributions.

This is what Harvard researchers Fiona Chen and James Stratton have found, by analyzing aggregated engineering data from Jellyfish, covering 300m work events and more than 700,000 employees at over 700 firms between 2021 and March 2026.

They had combined direct measurements of AI use with GitHub activity to estimate when organizations adopted AI assistants and agents, which can generate and submit code from prompts. The researchers’ analysis compared changes in firms’ output and workflows before and after adoption.

Evaluation process

After enterprises introduced coding agents, lines of code rose by 30% on average, commits by 20%, and pull requests by 23%, according to the data. But the rate at which Jira-tracked issues and larger feature requests were resolved did not change significantly.

The researchers also found no shift in the size or complexity of those tasks that could explain the result. Also:

  • Reviewing the added code took longer. The average time from submitting a pull request to merging it had increased 49% after agent adoption. The share of pull requests that reviewers sent back for changes had nearly doubled, while comments per pull request had risen by 35%. The share of employees doing code reviews had also increased by 14%.
  • The study found little evidence that agent adoption reduced staffing. Researchers compared workforce activity in Jellyfish data with LinkedIn information and could not attribute significant employment changes to AI.
  • Furthermore, automated review had not yet offset much of the added human workload. By March 2026, 80% of the firms in the data had used some form of AI code review, but AI agents accounted for only 23.3% of review comments and 10.8% of pull requests.

The results reflect a fast-changing field: 95% of firms in the study had adopted coding agents, but the data stops in March 2026, and teams may still be learning where the tools help most. Review burdens could ease with experience and improved systems.

For now, however, the study suggests that coding speed gains are being absorbed by longer, more demanding human review — not yet producing measurably more completed software.