AI coding agents produce more code, but not more software

Recently, it has been noted that AI-based coding agents can generate large volumes of functional code. However, developers using these tools warn that the accuracy of the generated code is not always reliable, which necessitates significant effort in reviewing the output. This phenomenon has been studied across various firms, revealing that human code review becomes a 'bottleneck' that hampers the overall efficiency of AI coding tools.

## Why it matters This finding is crucial as it calls into question the effectiveness of artificial intelligence in software production. Despite the capability of AI agents to generate code, the expected increase in software production has not materialized, suggesting that companies may not be achieving the anticipated return on investment when implementing these technologies.

## What we know The study's authors found that while the coding phase may experience an increase in efficiency, this benefit is counterbalanced by constraints in downstream production stages. The length of code review increases, and the likelihood of pull requests being delayed also rises, affecting overall productivity.

## What remains unclear It is still unclear how companies can optimize the code review process to fully leverage AI tools. The lack of evidence regarding increased software production or reduced employment following the adoption of these technologies raises questions about their large-scale implementation.