Study: AI Coding Agents Lengthen Review Times Despite Faster Output

Because agent-generated contributions require substantially more feedback and revision, organizations have had to shift human resources to manage the influx. The study documented a 14 percent increase in the share of workers actively conducting code reviews following the rollout of AI agents.

Interestingly, this shift has not translated into broader labor displacement. By examining active worker counts across Jellyfish and cross-referencing those figures with LinkedIn profiles across the analyzed firms, the researchers determined that they "cannot attribute significant employment changes to AI." Rather than replacing developers, the technology has changed how existing engineers spend their hours, redirecting significant human attention toward validating agent-generated output.

Limited Relief From Automated Code Reviews

To offset this surge in review demands, many organizations have explored using AI to assist in reviewing code as well. However, Chen and Stratton found that AI-assisted reviews have provided only marginal benefits so far.

Adoption rates across the studied companies are high: by the study’s March 2026 cutoff, 95 percent of firms had integrated AI coding agents, and 80 percent were utilizing some form of AI code review. Despite this widespread rollout, automated systems have taken on only a fraction of the actual reviewing workload. The data shows that AI agents were responsible for only 23.3 percent of all review comments and just 10.8 percent of all pull requests. Human engineers continue to handle the overwhelming majority of evaluation and decision-making.

This leaves engineering leaders facing what the study characterizes as a double-edged sword: gains in raw coding output are often neutralized by the corresponding rise in human code review time and effort. Consequently, it raises questions about whether the substantial time and financial expense required to implement AI coding agents produces a worthwhile return for most firms.

What it means for developers

For individual developers and technical leads, these metrics underscore the need to balance drafting speed against the total lifecycle cost of code. Faster generation can quickly become counterproductive if it floods the repository with pull requests that fail quality checks or demand extensive comment threads.

Software teams are still working through an active learning curve regarding when and how to deploy agents effectively. While upgrades and updates to agent outputs have continued past the March 2026 data cutoff, finding the right model for specific tasks remains critical to avoid overloading reviewers. Developers looking to experiment with different agentic systems can try top AI models cheaply through one API at https://apixoai.online to assess code quality and formatting before deploying them broadly across team repositories.

Moving forward, engineering organizations will need more practical experience to optimize the trade-offs between coding time and review time. Until agents learn to produce code that reliably meets review standards on the first attempt, the ultimate pace of software delivery will remain dictated by the human engineers tasked with checking their work.

Source: AI coding agents generate more code, but not more software — Ars Technica AI. Written by the Apixo team from that report.