· via Hacker News – Front Page (native)
Harvard working paper finds AI coding agents add 30% more code and 49% slower reviews
A Harvard working paper covering 718 firms found engineers using AI coding agents wrote about 30% more code, yet finished work did not measurably rise and code review times grew 49%.

Engineers at companies that adopted AI coding agents wrote roughly 30 percent more code, but finished work did not measurably increase and getting changes through code review took 49 percent longer, according to a working paper by two Harvard Ph.D. candidates reported by ComplexDiscovery.
The paper, by Fiona Chen and James Stratton, follows 725,938 workers across 718 companies and measures what happens at the human checkpoint once drafting speeds up. Its central finding is that the bottleneck in software delivery moved rather than disappeared.
What the study measured
Chen and Stratton built the study on data from Jellyfish, a company that sells an analytics platform tracking engineering activity, joined to LinkedIn employment records collected by Revelio Labs. The sample runs from January 2021 to March 2026 and covers client firms that consented to research use of their data, spanning 68.3 million Jira issues, 200.9 million commits, 49.1 million pull requests and 21.3 million calendar events, though only 213 of the 718 firms supplied calendar data.
The design is a staggered difference-in-differences comparison, tracking month by month how outcomes changed at adopting firms against firms that had not yet adopted. The authors argue adoption timing depended mainly on when firms learned about the tools and how long procurement took, with one interviewee's firm approving a tool in about a day and another citing six to 12 months on average. They acknowledge that larger firms adopted earlier and control for firm size, and they argue any remaining hidden difference would more likely flatter AI than undersell it, which would make the flat output result generous to AI.
Code volume up, finished work 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, the bundles of code submitted for review, rose 23 percent. The authors note these figures capture the added effect of agents at firms that already used AI coding assistants, not the effect of AI from a standing start.
Output is where the gain stops. The paper counts finished work as Jira issues resolved, with an issue counted complete once its code is deployed to production. Agents moved that measure by 0.12 issues per worker-month on a 3.67 baseline, with a standard error of 0.17, statistically indistinguishable from zero. The 95 percent confidence interval rules out an increase above about 12 percent of the baseline.
The authors tested the obvious counterargument, that engineers had shifted to larger tasks that a flat issue count would understate. They labeled a 0.5 percent sample of issues by size using GPT-4o, extended the labels across the dataset with a text-embedding model, and found no shift in average issue size. Pull requests did not grow larger either.
The bottleneck moved to review
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 drawing a formal change request rose about 12 percentage points from a 13 percent baseline, nearly doubling, and 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, meaning labor shifted toward reviewing.
AI review tools did not clear the queue. By March 2026, nearly 80 percent of the firms had used one, yet AI generated only 23.3 percent of review comments and only 10.8 percent of pull requests received even a single AI comment. The authors write that the bottleneck persists after firms adopt such tools. Interviews explain the mechanics: one junior engineer described a project where only a single senior engineer could approve changes, and only once a week, and said outages caused by AI-generated commits prompted new rules requiring senior approval of AI-generated code, deepening the bottleneck.
A working paper, not a settled result
The current version is dated Aug. 4, 2026, following a first version dated Jan. 7, and Chen circulated it as her job-market paper. Nothing ComplexDiscovery reviewed indicates journal acceptance or completed peer review. Coverage listed on Chen's website includes The Wall Street Journal and The Economist, and an Oct. 12 eDiscovery Today post by Doug Austin relayed an Ars Technica report on the findings.
Why it matters
This is one of the larger datasets on AI coding agents operating inside real engineering organizations, and its result complicates adoption metrics built on generation: more code written, statistically no more work shipped, and a review stage absorbing the difference. The authors' model separates two mechanisms, faster drafting queuing onto a fixed pool of reviewers and rising change requests and comments pointing to lower-quality drafts, which gives teams a concrete way to tell which problem they have. The lesson generalizes beyond software: wherever AI drafts and a human approves, approved output and reviewer time, not drafting volume, will show whether the tool is actually working.
- #ai
- #code-review
- #developer-productivity
- #software-engineering
- #research