Harvard study: coding agents add 23% more pull requests but no more finished work

Two Harvard economists, Fiona Chen and James Stratton, studied 300 million work events from 718 firms using the Jellyfish engineering analytics platform, from January 2021 to March 2026. They compared firms before and after they adopted AI coding assistants and agents. After adopting coding agents, firms produced about 30% more lines of code, 20% more commits and 23% more pull requests. But the number of finished Jira issues and epics did not rise by a statistically significant amount. The reason looks like review: the time from pull request to merge grew by 49% (3.45 extra days on a 7-day baseline), the share of pull requests sent back for changes nearly doubled, and comments per pull request rose 35%. The authors found no clear change in employment. It is a working paper using proprietary data, and agents have improved since the March cutoff.