AI NewsProductivityReported
Harvard paper finds AI agents move the slow step to code review
A new Harvard working paper tracks 655 firms adopting AI coding tools. The code review queue is where the productivity gains stop.

Why it mattersReview capacity is now the slow step. A team that adds coding agents without adding review capacity writes more code and finishes the same number of features.
The code a team writes is not the slow step anymore. A new Harvard working paper watched 655 firms adopt AI coding tools. It found the slow step moved to the person reading the code.
Fiona Chen and James Stratton of Harvard University released "Artificial Intelligence in the Firm: Bottlenecks in Software Production" on November 14. The paper is Chen's job-market paper. The data covers 445 million work events for 720,542 workers across 655 firms. It comes from the Jellyfish engineering analytics platform. The window is January 2021 to August 2026.
What they measured
The authors split AI coding tools into two generations. Assistants started in late 2022 and suggest local edits. Agents started in late 2024 and finish longer tasks on their own. The team used a staggered difference-in-differences design, comparing earlier and later adopters at each firm.
Twelve months after adoption, AI assistants raised lines of code by 30.4 percent. Commits rose 16.8 percent. Pull requests rose 20.9 percent. AI agents moved the same measures by 44.0 percent, 36.0 percent and 35.7 percent.
Finished work did not keep pace. With assistants, Jira issues resolved fell 1.5 percent and epics rose 3.4 percent, both near zero. With agents, issues resolved rose 13.7 percent and epics rose 18.0 percent. The authors call that borderline significant. Employment did not change in a way they could attribute to AI either.
Where the delay sits
After firms adopted AI agents, the time from a code change being opened to being approved rose 56.1 percent. The backlog of changes waiting for a reviewer rose 59.9 percent. Time once active review began rose 58.5 percent. The share of changes getting at least one review comment rose 13.1 percent. The number of comments per change rose 38.6 percent.
The authors say 79.1 percent of firms in the data have already adopted AI code review tools. The queue still shows up. Humans still write most of the review comments.
Chen and Stratton also note a structural reason the queue does not clear. Workers only review code they already know. In the data, 82.8 percent of real reviews match this pattern. Random assignment would give 17.2 percent. Code written by senior engineers is reviewed by other senior engineers 75.7 percent of the time. A firm cannot hire a reviewer at the same speed it hires a coder.
When the authors fit their model to the data, coding productivity rose 73.7 percent with AI agents. Review productivity fell 7.0 percent. Firms moved people. Coding labor fell 22.7 percent. Review labor rose 43.4 percent.
A team that adds AI agents without adding review capacity ends up with more code in a queue. It finishes the same number of features. The authors flag this directly. They simulate growth if the review queue stays slow, and the growth is smaller.
The paper is a working paper and has not been peer reviewed. Jellyfish reviewed the manuscript to check it did not disclose proprietary information. Jellyfish did not review the scientific content. The numbers above are the authors' own.
Source
Chen and Stratton, "Artificial Intelligence in the Firm: Bottlenecks in Software Production," Harvard University. Jellyfish is cited as the data partner.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


