AI Coding Agents Generate More Code, But Not More Software

A recent study of over 700 software companies found that AI coding agents boost the volume of code developers generate, but do not lead to a higher quantity of software released by those firms. According to Ars Technica, the research, which examined actual coding practices across numerous corporations, discovered that human code review emerges as a critical constraint on the overall efficiency of AI coding tools.

Consequently, there is little evidence to suggest that these firms see an increase in software output or a reduction in employment due to their utilization.

The study authors discovered that the introduction of AI coding agents at a firm results in a 30% increase in total lines of code produced, a 20% rise in the number of total commits, and a 23% boost in average pull requests. Nevertheless, all that additional code does not convert directly into improved software output at the firm level.

In fact, the resolution rate for issues and epics (i.e., comprehensive software features) tracked by tools such as Jira remained unchanged following the introduction of AI tools. The study also uncovered no substantial shift in the size or complexity of Jira-tracked issues after AI agents were introduced.

This discrepancy can be traced back to the code review process, which takes significantly more time on average after the implementation of AI coding agents. The average review process duration, from a pull request being submitted to it being merged into the codebase, increased by 49% on average after AI agents were introduced. This effect is evident in more detailed data, with the percentage of pull requests containing modifications nearly doubling, and the number of comments per pull request increasing by 35% after the shift to AI agents.

In response to these changes, the researchers found a 14% increase in the proportion of workers performing code reviews after the introduction of AI agents. Additionally, they noted that no significant employment changes could be attributed to AI after analyzing total active workers across Jellyfish and cross-referencing with LinkedIn data at those companies.

While AI could potentially assist in this review process, the researchers found that its impact has been marginal so far. Although 80% of measured firms employed some form of AI code review by March 2026, AI agents were responsible for only 23.3% of all review comments and 10.8% of all pull requests, indicating that humans still handle the majority of this work.