Study snapshot: Cursor’s agent in the wild
Suproteem Sarkar of the University of Chicago analyzed Cursor’s AI coding agent across tens of thousands of users to measure how agents change developer behavior and organizational throughput. The paper examines two core signals—how often developers call the agent and how often its edits are accepted—and compares productivity proxies before and after the agent became default in eligible organizations.
What stood out
The most striking quantitative result is a 39% increase in merged PRs at organizations that switched the agent to the default mode, measured relative to a baseline group. That signal sits at the center of the study’s productivity narrative: more merged work without clear increases in revert rates or larger code churn.
On the behavior side, the study finds that more experienced developers accept agent-generated code at higher rates, with roughly a 6% increase in acceptance per standard deviation of experience. The analysis suggests senior engineers also spend more time planning prompts and managing context, which correlates with higher-quality interactions with the agent.
Behavioral patterns and intent
Requests sent to the agent cluster into a few high-level intents. A majority of initial interactions were for implementation tasks, while other common intents included explaining code or errors and planning next steps. The distribution of these intents, coupled with acceptance behavior, paints a picture of agents being used both as code generators and as cognitive scaffolding for more senior engineers.
Why that matters
The combination of increased merged PRs and higher acceptance among experienced contributors points toward changes in workflow adoption and evaluation practices, rather than simply a volume effect. The lack of significant shifts in revert rates or average files touched per PR suggests the agent’s influence manifests in how work is completed, not necessarily in larger, riskier edits.
For a fuller read and the full dataset and methodology, see the original paper and analysis at Cursor’s blog and the linked study.
Original source: [https://cursor.com/blog/productivity](https://cursor.com/blog/productivity
