AI coding agents generate more code, but not more software
Ars Technica Kyle Orland ● Covered by 2 sources
AI coding tools crank out more code, but a big study says software teams don’t ship more. Human review slows everything down and eats the gains.
Based on reporting by Ars Technica, Kyle Orland — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI coding agents can spit out a lot of working-looking code. The catch is that programmers still have to check it, and that check is where the speedup starts to disappear.
A new study of coding work across hundreds of firms says there’s “little evidence” that AI tools have raised software output or cut jobs. The researchers found that whatever time gets saved while code is being written is largely swallowed later, when the work hits review and revision.
The paper comes from Harvard researchers Fiona Chen and James Stratton, who used aggregated analytics from Jellyfish, a tool that tracks engineering activity in detail. Their dataset covers 300 million individual work events, including commits and pull requests, plus issue-management data from more than 700,000 employees at over 700 software firms. The span runs from 2021 through March of 2026.
And the bottleneck is not subtle. According to the study, code review gets longer, pull requests are more likely to come back with revisions, and reviewers leave more comments when AI is in the mix. That means the gains from faster coding are being absorbed by the rest of the production process instead of turning into more shipped software.
The study’s basic message is almost painfully practical: generating code is not the same thing as delivering software. AI can flood the pipeline, but the pipeline still has to be cleaned up by humans.
My take — AI-written commentary, not fact-checked reporting
This is the part of the AI boom that gets skipped in the hype reels: code is cheap, trust is expensive. The machine can type faster than any team can review, which is a neat trick if the goal is to create more pull requests and a less cheerful Monday.
Read more about this at: Ars Technica