AI coding agents bottleneck: review is now the holdup

A new Harvard working paper on AI coding tools has put numbers on a frustration many engineering teams already feel. Economists Fiona Chen and James Stratton tracked software firms that adopted coding agents such as Claude Code, Cursor and Devin, and found individual developers writing substantially more code while the amount of finished software barely budged. The result is an AI coding agents bottleneck at the review stage. Every extra pull request an agent fires off still has to be read, checked and approved by a person, and that step has swollen to swallow the gains, the researchers write in Artificial Intelligence in the Firm: Bottlenecks in Software Production.

More code, no more shipped work

Chen and Stratton worked with data from Jellyfish, an engineering analytics platform whose clients agreed to research use of their records. The sample covers 718 firms and more than seven hundred thousand workers, with roughly 300 million work events drawn from version control, issue trackers and AI tool APIs between January 2021 and March 2026. Using a design that compares firms adopting the tools at different points, the paper isolates the effect of AI assistants first, then adds the effect of full agents on top. Assistants such as Copilot lifted lines of code modestly. Agentic tools moved every activity measure by a wide margin. Because each firm's post-adoption record is measured against firms that had not yet adopted, industry-wide trends are filtered out of the comparison.

Once firms adopted agents, lines of code per worker rose 30 percent, commits 20 percent and pull requests 23 percent, each statistically significant. The paper translates the code figure into nearly 1,500 additional lines per worker each month. Yet the proxies for finished work, resolved Jira issues and completed epics, moved only slightly and not significantly, according to the paper. The authors' estimates rule out a gain larger than 12 percent of the baseline, a ceiling well below the individual-level spike. The developers got faster while the firm, measured by what it ships, did not. The headline figures describe an AI coding agents bottleneck more than an adoption failure.

Review time ballooned

Review absorbed the surge, and that is the AI coding agents bottleneck in action. Average time from pull request submission to merge rose 49 percent after agent adoption. The share of pull requests sent back for changes nearly doubled, and comments per pull request climbed 35 percent, while pull request size did not change significantly. The paper reports that the bottleneck held even at firms that had already adopted AI code review tools, because human reviewers kept the final sign-off. Agents themselves also wrote a growing slice of the review traffic, leaving people to check machine-written comments on machine-written code.

Work moved, it did not leave

Firms responded by pulling more people into review work, one way of absorbing the AI coding agents bottleneck. The share of employees doing code review grew 14 percent. Overall employment at the firms showed no measurable change, which the researchers describe as a precise zero, cutting against the narrative that agents are quietly replacing engineers. To explain the pattern, Chen and Stratton interviewed 21 developers, product managers and engineering leads. One junior engineer put it plainly: "People could make a huge number of commits quickly, but code review was still the bottleneck." A senior engineer said a tool like Claude Code can produce a usable draft in seconds for routine work but makes frequent mistakes on harder tasks, leaving more review time, not less.

The usual cautions apply. The paper is a job market paper, a working paper that has not been peer reviewed, and its data end in March, before the newest generation of agents arrived. Jellyfish's clients also skew toward firms that already track engineering data closely. Even so, coverage of the findings across several outlets points the same way: when the review step is fixed, output scales with the slowest human in the chain. By March, the paper reports, 80 percent of the measured firms used some form of AI code review and 95 percent had implemented AI coding agents.

The measure that matters, the paper argues, is shipped work, not lines written. The findings have already circulated widely among developers, with a thread on a popular programming forum drawing hundreds of comments from engineers who recognized the pattern in their own teams. It echoes earlier reporting on this site's productivity beat, which has tracked how AI keeps widening workloads instead of shrinking them: see the Jabra WorkLife productivity paradox write-up. One industry write-up and a second detailed breakdown walk through the same figures.