Study finds AI coding agents increase code volume, but not software output

The widespread adoption of AI coding agents has led to a significant increase in the volume of code being written, but a major new study reveals this has not translated into more software being shipped. According to research from Harvard University, human code review has become a critical bottleneck, absorbing efficiency gains and preventing the promised boost in overall development output.

What did the study find about AI coding agents?

Researchers Fiona Chen and James Stratton analyzed over 300 million work events across more than 700 software development firms from 2021 through March 2026, using data from engineering analytics platform Jellyfish. According to their findings, published by Ars Technica, the introduction of AI coding agents led to a clear surge in raw coding activity. Firms saw a 30 percent increase in total lines of code generated, a 20 percent rise in commits, and a 23 percent increase in pull requests on average.

Despite this surge, the study found "little evidence that firms increase software output or reduce employment." The resolution rate for tracked software features and issues showed no statistically significant change after AI tools were introduced. The efficiency gains in code generation were entirely offset elsewhere in the development pipeline.

Why isn't more code leading to more software?

The research points to the code review process as the primary bottleneck. After AI agents are introduced, the average time between a pull request being submitted and merged into the codebase balloons by 49 percent. The data shows the share of pull requests requiring changes nearly doubles, and the number of comments per pull request increases by 35 percent. This indicates human reviewers are spending substantially more time scrutinizing and correcting AI-generated code, negating the time saved during initial creation.

How are development teams responding?

In response to the increased review burden, the study observed a 14 percent increase in the share of workers performing code reviews. While 80 percent of firms used some form of AI-assisted code review by March 2026, AI agents were responsible for only 23.3 percent of review comments and 10.8 percent of pull request reviews, indicating humans still handle the vast majority of this critical quality assurance work. The researchers found no significant changes in overall employment levels attributable to AI adoption.

What does this mean for developers using AI tools?

The study, detailed in the original Ars Technica article, presents a reality check for teams expecting AI to be a straightforward productivity multiplier. It suggests that simply generating more code faster does not accelerate software delivery if downstream processes like review and integration cannot keep pace. The findings highlight the importance of optimizing the entire development workflow, not just the coding phase, and of setting realistic expectations for AI's impact on project timelines and team structure.

How to explore AI-assisted coding in Grafy

For developers and teams looking to experiment with AI coding while managing the review bottleneck, Grafy Code provides a controlled environment. As a multi-model AI coding studio, it allows you to generate, edit, and run code with various cloud models side-by-side. Crucially, its integration with Grafy Graph means every prompt, code output, and subsequent edit is visually tracked in a branching graph. This creates a transparent, auditable history of AI-generated code, which can significantly streamline the human review process by providing clear context for how and why code was written. You can learn more about this approach in our blog post on AI workflow without exports.

Teams can use this structured environment to test prompts, compare model outputs for a given task, and refine their AI coding practices before integrating into larger projects. For broader research on AI development trends, you can also use Grafy Reader to upload and interrogate academic papers or reports like this one, using AI to summarize findings and extract key data.

Frequently asked questions

Did the study find AI coding tools reduced developer jobs?

No. The Harvard researchers found they "cannot attribute significant employment changes to AI" after analyzing data from Jellyfish and LinkedIn. While the share of workers doing code reviews increased, overall headcount at the studied firms did not show a statistically significant shift.

Are AI coding assistants or agents better for productivity?

The study differentiated between assistants (auto-completing human code) and agents (writing code autonomously). It found the significant review bottleneck was specifically linked to the introduction of AI agents, which generate larger volumes of code that require substantial human verification and correction.

Can AI help review the code it generates?

Partially. While 80% of firms used AI-assisted review tools, AI was only responsible for about a quarter of review comments and 11% of pull request reviews by March 2026. Human reviewers still perform the majority of this work, suggesting AI is not yet a complete solution for the bottleneck it helps create.

---

Source: Hacker News — front page — AI coding agents generate more code, but not more software