Harvard researchers Fiona Chen and James Stratton analyzed aggregated metrics from Jellyfish, a platform that records detailed engineering activity. Their dataset spans 300 million work events,including commits, pull requests, and issue-tracking records,across more than 700,000 staff at over 700 software firms between 2021 and March 2026. The study compared periods before and after the deployment of AI-driven coding assistants and autonomous coding agents.
When autonomous coding agents entered a company’s workflow, the researchers observed a 30 percent rise in total lines of code, a 20 percent increase in commit count, and a 23 percent uplift in pull-request volume. These figures represent average changes across the sampled organizations and illustrate that AI agents can substantially accelerate raw code production.
Despite the surge in code output, the study found no statistically significant shift in the rate at which issues or epics,large software features tracked in tools such as Jira,were resolved. The size and complexity of these tracked items also remained stable, indicating that faster code generation did not translate into faster feature delivery.
The researchers linked the disconnect to a lengthening of the code-review stage. After AI agents were adopted, the average interval between a pull request’s submission and its merge grew by roughly 49 percent. The proportion of pull requests that received change requests nearly doubled, and the number of comments per request rose by about 35 percent, according to the authors.
A modest rise in the share of staff performing reviews accompanied the longer cycles, with a 14 percent increase observed after AI agents were introduced. However, the authors note that total employment levels remained unchanged, and they could not attribute any significant workforce shifts directly to AI adoption.
By March 2026, 80 percent of the firms surveyed employed some form of AI-assisted code review, yet autonomous agents contributed only 23.3 percent of review comments and 10.8 percent of pull requests. Overall, 95 percent of the sampled companies had deployed AI coding agents, suggesting widespread adoption despite the pending learning curve and the current trade-off between faster coding and slower review.