What happened
Engineers at companies that adopted AI coding agents produced roughly 30 percent more code with no measurable increase in completed work, according to a working paper by Harvard University Ph.D. candidates Fiona Chen and James Stratton, ComplexDiscovery reported. The study spans 725,938 workers at 718 firms between January 2021 and March 2026, combining engineering activity data from Jellyfish with LinkedIn employment records compiled by Revelio Labs.
Per worker-month, added code grew by 1,494 lines from a base of 5,017. Commits rose 20 percent and pull requests 23 percent. Finished work, which the paper measures as Jira issues resolved once code reaches production, moved by 0.12 on a 3.67 baseline. That change is statistically indistinguishable from zero, and the authors' confidence interval excludes any gain above roughly 12 percent.
The slowdown showed up in review. The time between submitting a pull request and merging it lengthened by 3.45 days from 7.03, a 49 percent rise. The share of pull requests drawing a formal change request climbed about 12 percentage points from 13 percent, and comments per pull request increased 35 percent. Nearly 80 percent of the firms had used an AI review tool by March 2026, yet such tools wrote just 23.3 percent of review comments.
Context
The paper is Chen's job-market paper. Its current version is dated Aug. 4, 2026, and it shows no sign of completed peer review. The authors attribute differences in adoption timing mainly to awareness and procurement delays, and they estimate that any drop in overall employment was no larger than 2.9 percent. Harvey and Cooley announced legal AI tools on Oct. 7 and Oct. 8 that each keep a lawyer at a review step.
The briefing
The week in tech, in one email.
The stories worth your time, every week. Unsubscribe anytime.
