
Harvard Study Finds AI Coding Agents Shift Bottleneck to Code Review
Harvard researchers found that AI coding agents increased code output by 30 percent but did not raise finished work, while code review time grew 49 percent across 718 firms studied.
A Harvard working paper by Ph.D. candidates Fiona Chen and James Stratton shows that firms adopting AI coding agents saw engineers write about 30 percent more code, yet completed work did not increase measurably and code review took 49 percent longer. The study analyzed data from 718 companies covering 725,938 workers between January 2021 and March 2026.
Increased Output Without Delivery Gains
Lines of code added rose by 1,494 per worker-month against a baseline of 5,017, a 30 percent increase. Commits rose 20 percent and pull requests rose 23 percent. However, finished work measured as Jira issues resolved changed by only 0.12 issues per worker-month on a 3.67 baseline, a result statistically indistinguishable from zero. The 95 percent confidence interval rules out an increase above about 12 percent of the baseline. Sampling issue sizes with GPT-4o found no shift to larger tasks, and pull requests did not grow larger either.
Review Queue Becomes Primary Bottleneck
The 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. Interviews showed that after outages caused by AI-generated commits, some firms began requiring senior engineer approval for such code, adding further bottlenecks.
Legal AI Tools Face Similar Checkpoint Dynamics
Two legal AI announcements in October 2026 follow a comparable design. Harvey's motion-to-dismiss workflow, co-developed with LexisNexis, produces a first draft in hours but keeps human checkpoints at moments where a lawyer's judgment matters most. The announcement does not state how long lawyers spend at those checkpoints. Cooley's redaction agent, which the firm is still developing with Google Cloud, leaves redaction decisions to attorneys at the last step before a filing goes to court, and the announcement carries no error rate. Under Federal Rule of Civil Procedure 11(b), an attorney presenting a motion certifies its legal and factual contentions, an inquiry that stays with the attorney regardless of drafting speed. The Harvard study suggests buyers should measure hours from upload to approved work product, the share of drafts sent back, and reviewer time per matter, keeping in mind that the study examined software, not legal work.
Employment Effects Remain Limited
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 that these are short-run effects in partial equilibrium, which could mask job losses at competing firms. A separate National Bureau of Economic Research working paper revised in September 2026 found a similar pattern using data on over 500,000 GitHub developers, with a 240 percent cumulative rise in commits from autonomous agents shrinking to 30 percent for actual releases. The Harvard paper's current version is dated August 4, 2026, and is a job-market paper with no indication of completed peer review.
Sources: complexdiscovery.com
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.