AI Coding Agents Write More Code, but Review Keeps Software From Shipping Faster

AI coding agents can produce code far faster than developers can write it by hand, but a new study suggests that extra output is not turning into proportionally more finished software. According to the research, the efficiency gains are being soaked up downstream, mostly at the stage where people have to check the work.

The choke point is review. Every AI-generated change still has to be read, tested, and approved by someone with limited time. When the volume of proposed code jumps, the queue of pull requests grows. Reviewers either slow down or start skimming, and neither outcome lets a team ship features at the pace the raw output numbers imply.

The finding matters for engineering managers who are buying these tools on the promise of faster delivery. Metrics such as lines of code generated or tasks opened can make a team look more productive than it is. The question that counts is how quickly working software reaches users, and on that measure the gains look smaller.

The takeaway points less at the quality of the models and more at the workflow around them. Smaller, well-scoped changes are easier to review. Stronger automated tests can catch errors before a human ever looks at the code. Review practices that scale with the amount of generated code, rather than staying fixed, may matter as much as the agent itself.

For now, the study is a reminder that writing code was never the only hard part of building software. As agents take over more of the typing, the bottleneck is moving to judgment, verification, and the people who have to sign off.