The productivity impact of coding agents (cursor.com)

A University of Chicago study by Suproteem Sarkar analyzed tens of thousands of Cursor users to quantify early productivity effects from making Cursor’s coding agent the default. Comparing organizations already using Cursor (“eligible”) with a baseline group, the study reports a 39% relative increase in merged PRs after the agent became default, while PR revert rates stayed flat, bugfix rates slightly decreased, and average lines/files changed per merged PR were unchanged. The researchers tracked two behavioral signals—how often users query the agent and how often they accept its code edits—and sampled 1,000 users’ conversation starts to classify intents: ~61% asked the agent to implement code, with the remainder focused on explanations/errors and planning. Technically significant patterns emerged around experience and workflow: more experienced developers write more plans before coding and accept agent suggestions more often—each standard-deviation increase in years of experience corresponds to about a 6% rise in acceptance rate—contrary to expectations that juniors would lean on agents more. Possible reasons include seniors’ better context management, use of custom rules, greater confidence evaluating generated changes, and working on more well-scoped tasks. The study highlights practical implications for agent design (support planning, alignment, and review), adoption dynamics across seniority, and the challenge of defining robust economic metrics for AI’s impact on software engineering.