
Sign up to save your podcasts
Or


This National Bureau of Economic Research paper investigates the productivity impact of generative AI on software development by tracking three generations of coding tools. The authors analyse data from over 100,000 GitHub developers to compare simple autocomplete features, interactive sync agents, and autonomous async agents. While these tools dramatically increase task-level activity, such as code volume and commit frequency, the study reveals that these gains diminish significantly as they move toward final output. This attenuation is attributed to a "weak-link" effect, where human bottlenecks in code review and project management limit the overall speed of software releases. Finally, the research finds that while AI has increased the supply of new applications in digital marketplaces, there has been no corresponding increase in total user engagement.
By HimanshuThis National Bureau of Economic Research paper investigates the productivity impact of generative AI on software development by tracking three generations of coding tools. The authors analyse data from over 100,000 GitHub developers to compare simple autocomplete features, interactive sync agents, and autonomous async agents. While these tools dramatically increase task-level activity, such as code volume and commit frequency, the study reveals that these gains diminish significantly as they move toward final output. This attenuation is attributed to a "weak-link" effect, where human bottlenecks in code review and project management limit the overall speed of software releases. Finally, the research finds that while AI has increased the supply of new applications in digital marketplaces, there has been no corresponding increase in total user engagement.