AI coding agents generate more code, but not more software
A Harvard working paper finds that greater coding activity did not translate into a statistically significant rise in completed work.
TL;DR
- The working paper analyzed about 300 million engineering events across 718 firms and more than 725,000 workers.
- Coding-agent adoption was associated with more code, commits and pull requests, but no statistically significant rise in completed issues or larger features.
- The measured review interval also grew; the study is observational and does not establish a universal causal effect.
Harvard researchers Fiona Chen and James Stratton analyzed about 300 million engineering work events across 718 firms and more than 725,000 workers, covering January 2021 through March 2026. Their August working-paper version studies AI coding tools. [1]
The authors estimate that coding-agent adoption was associated with 30% more lines of code, 20% more commits and 23% more pull requests per active worker. They found no statistically significant increase in resolved issues or larger software features in the measures they analyzed. [1]
Ars Technica reported that the interval from pull-request submission to merge rose by an average of 49% after firms introduced coding agents. This is elapsed process time, which can include waiting and rework, rather than a direct measure of reviewer labor. [2]
The results are observational and come from firms using an engineering analytics platform, so they do not prove that AI agents caused the changes in every organization. The estimate for completed work is uncertain, not proof of exactly zero effect. [1]
Why it matters
The study shows how review and integration can limit the conversion of increased AI-generated code into completed software.
Editor's note
This is an observational estimate, not a universal causal result; code volume and completed work are distinguished.