AI agents boost coding output as human reviews slow software delivery
A study of software firms finds that AI coding agents increase the amount of code teams produce, but have not led to a measurable rise in completed software work or a reduction in employment. The researchers say the gains are absorbed by longer human review, which has become a bottleneck.
Drawing on 300 million work events involving more than 700,000 employees at over 700 firms from 2021 to March 2026, the study found that AI agents were associated with 30 per cent more lines of code, 20 per cent more commits and 23 per cent more pull requests. The average time from submitting a pull request to merging it rose by 49 per cent; requests for changes nearly doubled and comments per pull request rose by 35 per cent. Although 80 per cent of firms used some form of AI code review, its impact was marginal.
- AI agents produced more code and pull requests.
- Human review took 49 per cent longer.
- Completed software work and employment showed no significant change.
New here? Start with this
Software companies have been adopting AI coding agents – tools that automatically write computer code – in hopes of speeding up their development work and producing software faster. These tools have become increasingly common over the past few years.
Researchers examining over 700,000 software workers across more than 700 companies found that AI agents significantly boost code production. Teams using these tools produced about 30 per cent more code and generated substantially more work to be checked and approved.
However, this increase in output has not led to more software being released to users or to any reduction in employment. Instead, the extra code takes substantially longer for human colleagues to review and approve, suggesting that the time burden has simply shifted from writing code to checking it.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
This development represents healthy maturation of AI-assisted software development, where tools augment rather than replace human expertise. The increased time spent on human review is a feature of responsible integration—teams are thoroughly evaluating AI-generated code to enhance quality assurance. Employment remains stable, suggesting genuine augmentation of workforce capacity, whilst higher volumes of commits and pull requests indicate accelerated iteration and feature development. As teams become familiar with AI tools, review processes will improve, and the current friction reflects a natural learning curve.
The case against
The data presents a troubling pattern of substitution rather than genuine productivity gain. Although AI agents produce more code, this hasn't translated into faster software delivery—instead, human review time has increased 49 per cent, change requests have nearly doubled, and comments per request rose 35 per cent, suggesting AI-generated code requires extensive correction. This has effectively shifted labour from writing to reviewing rather than eliminating bottlenecks, complicating workflows without measurable improvements in completed work or overall efficiency.
AI Research Science Software Technology
Read the full article at the source →
Originally published by Ars Technica as “AI coding agents generate more code, but not more software”.