Harvard Study: AI Coding Agents Shift Bottleneck to Code Review, Not Output
Engineers at 718 firms wrote 30% more code with AI agents but finished work did not rise; pull request merge time increased 49% and change requests nearly doubled.

By Clara Tan, Artificial Intelligence
6 min read
Engineers at firms that adopted AI coding agents produced about 30 percent more code but finished work did not measurably increase, and the time to get code through review grew by nearly half, according to a Harvard working paper by PhD candidates Fiona Chen and James Stratton.
The study, based on data from 725,938 workers at 718 companies using Jellyfish engineering analytics and Revelio Labs employment records, covers January 2021 through March 2026. It uses a staggered difference-in-differences design comparing adopting firms against those that had not yet adopted.
Code up, output flat
Lines of code added rose by 1,494 per worker-month against a baseline of 5,017, about 30 percent. Commits rose 20 percent and pull requests 23 percent. But finished work, measured as Jira issues deployed to production, moved by just 0.12 issues per worker-month on a 3.67 baseline, with a standard error of 0.17. The 95 percent confidence interval rules out an increase above roughly 12 percent of baseline, well short of the 30 percent rise in code written.
The authors tested whether engineers shifted to larger tasks. A 0.5 percent sample of issues labelled by GPT-4o, extended via text embeddings, showed no shift in average issue size, and pull requests did not grow larger.
The review queue is where the gain stops
Time from pull request submission to merge rose 3.45 days on a 7.03-day baseline, a 49 percent increase. The share of pull requests receiving a formal change request increased by about 12 percentage points from a baseline of 13 percent, nearly doubling. Review comments per pull request rose 35 percent. The share of workers doing any code review rose 14 percent from a base of about 29 percent, while the share writing code showed no measurable change.
AI review tools did not remove the bottleneck. By March 2026, nearly 80 percent of firms had used one, yet AI generated only 23.3 percent of review comments, and only 10.8 percent of pull requests received even one AI comment. The authors write that the bottleneck persists after firms adopt such tools.
Interviews reveal the human side: one junior engineer described a project where only one senior engineer could approve changes and reviewed code only once a week. "After some outages that were caused by AI-generated commits, there were also new rules requiring senior engineer approval for AI-generated code, which created even more review bottlenecks," the engineer said.
Two channels, different fixes
The authors model two channels. First, faster drafting sends more code to a fixed pool of reviewers, a problem that moving labor into review can largely relieve. Second, the rise in change requests and comments suggests lower-quality drafts, which would require better technology. Buyers should care which channel dominates, because the two call for different spending.
No employment decline detected
The estimates rule out a decline in overall employment larger than 2.9 percent. A noisier estimate rules out a decline in engineering employment larger than 13.8 percent. The authors caution these are short-run effects in partial equilibrium.
Legal AI tools follow a similar pattern
Harvey's motion-to-dismiss workflow (announced 7 October) and Cooley's redaction agent with Google Cloud (8 October) both automate a first pass and keep a human at the checkpoint. Neither announcement reports how long lawyers spend at that checkpoint or error rates. Under Federal Rule of Civil Procedure 11(b), the attorney presenting a motion certifies it after reasonable inquiry, a duty software evidence suggests may absorb some drafting savings.
The paper also notes that AI authorship leaves metadata trails in collaboration systems, bot accounts, "Co-authored-by" lines in commits, that retention schedules and legal holds may need to reach.
The current version is dated 4 August 2026 and is a job-market paper with no indication of completed peer review. A separate NBER working paper revised September 2026 reports a similar falloff using data on over 500,000 GitHub developers.
Sources
- 1ComplexDiscoverycomplexdiscovery.com
- 2Harvard Working Paper (Chen & Stratton)scholar.harvard.edu