AI coding agents can help developers produce code faster, but does more code mean software work is completed sooner? A working paper by two Harvard doctoral candidates examines this question using company data. According to the OSChina summary, the sample includes 718 companies and 725,938 employees. The reported change is not limited to higher coding output: code review is also identified as a downstream bottleneck that teams need to address.
Key findings
- After AI coding agents were adopted, monthly added lines of code per employee rose from 5,017 to 6,511, an increase of about 30%.
- Commit counts rose 20%, and pull requests increased 23%. Both coding output and activity in the collaboration workflow went up.
- More code did not mean that tasks were automatically completed. The material says the paper also used Jira data to examine work progress, but the supplied summary cuts off before giving the relevant measures or results.
- The headline points to code review as the current constraint. In other words, AI may accelerate code generation without removing the work required to inspect, understand, and integrate the resulting changes.
These figures are a reminder not to judge coding AI solely by lines of code or commit counts. Those metrics describe development activity, not code quality, delivery speed, or whether a task has ultimately been completed. If code generation accelerates but review capacity does not keep pace, additional output can shift pressure to examining and processing pull requests. Teams adopting agents should therefore pay attention to review workflows and developer capacity, rather than focusing only on increasing the number of submissions.
The available material does not include the paper’s full methodology, comparison design, or Jira findings. It cannot establish from this summary alone that AI necessarily reduces overall efficiency, nor does it quantify any increase in review time. A more measured conclusion is that the reported growth in coding activity is not the same as proof that work is finished. The next challenge for AI-assisted development is not only faster generation, but also review and delivery processes capable of handling the extra output.
Source: OSChina
Comments
Checking sign-in status...
Loading comments...