Coding agents raised code output, not finished work

A working paper by Harvard's Fiona Chen and James Stratton studied how AI coding agents affect real software teams. They analyzed about 300 million work events across more than 700 firms from 2021 to March 2026, combining Jellyfish engineering telemetry with GitHub activity. It has not yet been peer reviewed.

After agents arrived, firms produced on average 30% more lines of code, 20% more commits and 23% more pull requests. Completed issues and epics did not rise in a statistically significant way.

Time from opening to merging a pull request rose 49%, the share of pull requests with changes requested nearly doubled, comments per pull request rose 35% and the share of staff assigned to review rose 14%. By March 2026, 80% of firms used AI review tools, but only 23% of review comments came from AI.

A commit records a code change, a pull request asks to merge changes into the main code, and code review is when another developer checks them for bugs and security issues before release.

Speeding up one step does not speed up the whole process; the bottleneck moves downstream. The same applies to any AI-accelerated work, from proposals to content: more output means little without capacity to check it and turn it into results.

Earlier research mostly used controlled experiments that timed a single task with and without AI and often found 20% to 50% speedups. The Harvard paper instead uses years of everyday data from real companies, where review, testing and coordination absorb much of the lab speedup, explaining the gap between the two kinds of results.

Find the next bottleneck before adopting an AI tool, measure finished outcomes rather than volume, do not cut review steps as the load grows, and when commissioning software, judge by delivery dates rather than coding speed.