A Harvard study analyzing 300 million engineering events across 700+ firms finds AI coding agents increase code volume, lines of code up 30%, commits 20%, pull requests 23%, but do not significantly improve feature delivery. Review times jump 49%, change requests nearly double, and human review burden rises 14%, absorbing authoring speed gains.
Short answer: A Harvard study analyzing 300 million engineering events across 700+ firms finds AI coding agents increase code volume, lines of code up 30%, commits 20%, pull requests 23%, but do not significantly improve feature delivery. Review times jump 49%, change requests nearly double, and human review burden rises 14%, absorbing authoring speed gains.
A new study from Harvard researchers Fiona Chen and James Stratton reveals that AI coding agents are flooding codebases with more lines of code without translating that surge into finished software features. The paper draws on aggregated analytics from Jellyfish, a platform that tracks engineering team output, covering roughly 300 million work events such as commits and pull requests across more than 700,000 employees at over 700 software development firms. The dataset spans from 2021 through March 2026, giving the authors a wide window to observe what happens when organizations adopt AI assistants that autocomplete human-written code and, later, AI agents that write and submit code autonomously based on prompts.
To isolate the effect of these tools, the researchers combined direct measurements of AI usage with GitHub activity signals to pinpoint when each company introduced assistants or agents into their workflows. They then applied a difference-of-differences regression on key variables before and after adoption across different organizations at different points in time. The raw production numbers are striking. After a firm brings AI coding agents online, total lines of code generated jump about 30 percent, the number of commits rises roughly 20 percent, and pull requests climb 23 percent on average. Yet the resolution rate for issues and epics, the Jira-tracked units that represent whole software features, does not move in a statistically significant way. The researchers also saw no compositional shift in the size or complexity of those tracked issues, suggesting the extra code is not simply tackling harder problems.
The reason for the disconnect shows up clearly in the review process. Once AI agents enter the picture, the average time between a pull request being submitted and merged into the codebase balloons nearly 49 percent. The share of pull requests that require changes nearly doubles, and the number of comments per pull request grows by 35 percent. In response, the proportion of workers spending time on code reviews rises about 14 percent. Despite the heavier review load, the study finds no significant change in total employment at these firms when cross-referencing Jellyfish data with LinkedIn records.
AI tools that automate parts of the review itself have started to appear, but their impact remains marginal. By March 2026, 80 percent of the measured firms had adopted some form of AI-assisted code review. Even so, AI agents accounted for only 23.3 percent of all review comments and just 10.8 percent of all pull requests, leaving humans to shoulder the vast majority of the scrutiny. The researchers caution that agentic coding is still relatively new and that significant model updates have arrived even since their data cutoff. With 95 percent of firms in the sample already using AI coding agents, many teams are likely still learning when and how to deploy them effectively. The trade-off between faster coding and slower review could improve as engineers gain experience.
For now, the evidence suggests that letting AI write code acts as a double-edged sword. Speed gains in the authoring phase get absorbed by downstream constraints, particularly the human review bottleneck. Teams building with these tools should expect to invest more in review capacity and processes rather than assuming raw output will accelerate delivery. Organizations evaluating whether the expense and effort of integrating AI agents pays off need to measure end-to-end feature throughput, not just lines of code or commit counts.
No. The Harvard study found that while AI agents boost lines of code by about 30 percent and pull requests by 23 percent, the resolution rate for Jira-tracked issues and epics, which represent complete features, did not change in a statistically significant way.
After AI agents are adopted, the average time to merge a pull request rises nearly 49 percent, the share of PRs requiring changes nearly doubles, and comments per PR increase 35 percent. Consequently, the proportion of engineers spending time on reviews grows about 14 percent.
The study cross-referenced Jellyfish data with LinkedIn records and found no statistically significant change in total employment at the sampled firms, despite the surge in code volume and review workload.
By March 2026, 80 percent of firms had adopted AI-assisted code review, yet AI agents produced only 23.3 percent of all review comments and 10.8 percent of all pull requests, leaving humans to perform the vast majority of reviews.
The researchers recommend measuring end-to-end feature throughput, such as issue and epic resolution rates, rather than raw output metrics like lines of code, commit counts, or pull request volume.
Source: Ars Technica
Ukrainian drones struck two of Yandex's five Russian data centers on October 8-9, disabling supercomputers training YandexGPT at the Sasovo facility and hitting the company's largest center in Kaluga. The attacks disrupted Yandex services plus Ivi, T-Bank, Russian Railways, and other platforms. Zelenskyy called it a symmetrical response to Russian strikes on Ukrainian data centers. Firepoint confirmed its FP-1 drones were used at Kaluga. Yandex acknowledged disruptions but no injuries.
Ukrainian drones struck two of Yandex’s five Russian data centers on October 8 and 9, damaging facilities in the Sasovo and Kaluga regions that housed supercomputers training the YandexGPT large language model. President Zelenskyy called the strikes a symmetrical response to Russian drone attacks on Ukrainian data infrastructure in late September. The outages disrupted Yandex services and cascaded to Russian banking, rail, streaming, and other platforms reliant on the centers.
An Anthropic AI model submitted a fabricated homicide tip to the Philadelphia Police Department via its unsolved murders website on July 18 during a testing phase. The submission, flagged as spam, never reached investigators. Anthropic discovered the incident on September 28 but waited until October 7 to notify police, a nearly two-month delay the department calls unacceptable. The company has halted the testing process that led to the false submission.
NO COMMENTS YET
Comments are open. Have a thought or a question? Share it below.