AI coding agents generate more code, but not more software

AI coding agents are generating more code than ever, but that increase isn't translating into more software delivered. A recent study reveals that the efficiency gains are absorbed by the human review bottleneck. In other words: we write faster, but we don't ship more products.

Context

The promise of AI in software development has been enormous: automate repetitive tasks, speed up code writing, and ultimately reduce delivery times. Tools like GitHub Copilot, Amazon CodeWhisperer, and other autonomous agents have proliferated, promising to revolutionize the industry. However, the reality is more complex: generated code still requires human oversight, and that's where the problem arises.

Details

The study, conducted by researchers from several universities and tech companies, analyzed development teams using AI agents to generate code. The results showed that while code production increased significantly (up to 50% in some cases), software delivery did not grow proportionally. The main reason: human reviewers can't process that extra volume of code without becoming a bottleneck. Additionally, the quality of generated code isn't always optimal, adding time for correction and validation. The data indicates that for every hour saved in writing, 45 minutes are lost in additional review.

Analysis

This finding highlights that AI is not a silver bullet. While AI agents are useful for specific tasks, their impact on overall productivity is limited by human processes that don't scale at the same pace. The winners are companies that manage to redesign their workflows to integrate AI effectively, for example, by also automating review or setting stricter quality metrics. The losers: those who expect AI alone to solve all problems. My take: the key is orchestration, not generation.

Implications

In the short term, we'll see a renewed focus on tools that automate code review and error detection, such as smart linters or AI-based automated testing systems. In the long term, the industry could evolve toward a model where AI agents not only generate code but also validate and deploy it, reducing human intervention. But to do so, trust and regulatory barriers must be overcome. Meanwhile, companies must be aware that productivity isn't measured in lines of code, but in value delivered.

Next time you see an AI agent spit out hundreds of lines, ask yourself: how many will actually make it to production?

Source: Ars Technica