AI Coding Agents Lift Code Output 30% but Review Time Jumps 49%

More code, about the same amount of finished software

AI coding agents produce more code, but the software isn't shipping any faster. A Harvard working paper by Fiona Chen and James Stratton reaches that conclusion, and it has quickly become a reference point in the argument over whether tools like Claude Code, Cursor and GitHub Copilot pay off at the company level. The researchers used aggregated data from the engineering-analytics platform Jellyfish: about 300 million work events, such as commits and pull requests, from more than 700,000 employees at over 700 software firms, covering 2021 through March 2026.1 Other coverage puts the sample at 718 firms and says the data included HR records, version control, issue trackers, calendars and generative-AI tool usage from GitHub Copilot, Cursor and Claude Code.2

On the coding side, the gains are clear. When firms adopted AI coding agents, total lines of code rose about 30%, commits about 20% and pull requests about 23%.1 In absolute terms, that is roughly 1,494 extra lines per worker-month on a baseline of about 5,017.23 Completed work tells a different story. The study's measure of shipped software is resolved Jira issues and completed epics, and the rate at which those were resolved did not rise in a statistically significant way.8

The paper also separates autocomplete-style coding assistants from agents. The assistants produced smaller gains: 12% for lines of code, 9% for commits and 5% for pull requests, and only the rise in commits was statistically significant.2

Code review is where the gains disappear

The authors say the missing productivity is "absorbed by downstream constraints," and point mainly to code review.1 After firms adopted agents, the average time from opening a pull request to merging it rose 49%.1 That is 3.45 extra days on a 7.03-day baseline.23 The share of pull requests that received a formal change request went from about 13% to roughly 25%, close to double.28 Firms reacted by moving people: the share of workers doing code review grew 14%, while the share writing code stayed about the same.28

The most important detail may be that pull requests did not get bigger. Average issue size and PR size stayed roughly constant, so the authors rule out the obvious explanation that agents simply produce larger diffs that take longer to read.23 The paper's model treats the near-doubling of change requests as a possible sign that first drafts are lower quality.23 That puts the problem in how much rework each change needs, not only in how much code is arriving.

The study also undercuts claims about layoffs. The researchers could not link significant employment changes to AI after checking active worker counts against LinkedIn data.1 One commentator who read the paper says it rules out employment declines of more than about 3%.9

Where the coverage disagrees

The overall story is consistent across outlets, but a few numbers differ, and readers comparing articles should know which is which. Ars Technica and ComplexDiscovery both say comments per pull request rose 35%.123 InfoQ-derived summaries give 53%.23 The 35% figure matches the wording quoted from the paper and is the better-supported number.6

Coverage also differs on resolved issues. Some summaries say agents produced a 12% increase in resolved issues.2 A closer reading presents the estimate as statistically insignificant, with about 12% as the upper bound.28 The second reading is the right one. The study does not show that finished work went up by 12%. It shows that any increase was too small to separate from zero, and 12% is roughly the ceiling.

Some secondary accounts go further than the paper. One describes a single organization that saw a 40% rise in review hours against a 5% rise in deployed features.5 None of the other reports mention that example, and it should be treated with caution. The reaction on Reddit, where some readers concluded the paper found "no productivity" from Claude Code, was also wrong. Commenters quoting the abstract pointed out that it says both kinds of tool increase coding productivity, but the gains do not fully reach output.10

There are real limitations. This is an observational job-market paper with no sign that it has finished peer review.23 Its data ends in March 2026, before the agent releases of recent months.26 It also contradicts Jellyfish's own earlier messaging: a March press release said AI was speeding up coding and that code quality did not appear to be suffering.23

Other data shows the same review bottleneck

The finding is not an isolated result. It fits a set of 2026 industry figures, and they point the same way. A Carnegie Mellon and Stanford study of one AI-forward company found pull-request volume grew 3.1x while the pool of reviewers grew only 1.5x. Over the same period, the share of PRs that received any written human feedback fell from 39% to 21%.27 LinearB's benchmarks, covering 8.1 million pull requests, show a 30-day merge rate of 32.7% for AI-assisted PRs versus 84.5% for unassisted ones, with reviewer pickup times measured in hours rather than minutes.24 CircleCI's 2026 delivery data shows throughput on feature branches up 59% year over year, while median main-branch throughput fell.22

The volume is also growing fast. Anthropic says its typical engineer merged eight times as much code per day in Q2 2026 as in 2024, and The Pragmatic Engineer reported that agent-authored pull requests on GitHub outnumbered human-authored ones in August 2026.27 Harvard's sample ended before both of those data points, so the bottleneck it describes has most likely grown since.

There is a quality concern too. Industry write-ups of GitHub's review telemetry say agent-authored PRs carry more technical debt on average, yet human reviewers approve them more readily.24 Taken together, those findings describe review queues that are getting longer while each review gets less careful.

The vendors are moving into review

The paper also tested the obvious fix, which is to have AI review AI-written code. By March 2026, nearly 80% of the firms studied had used some AI code-review tool. Even so, AI wrote only 23.3% of review comments, and only 10.8% of pull requests received any AI comment at all.23 The authors conclude that the bottleneck remains after firms adopt these tools.23

The major vendors have made review their main focus since then. Anthropic made Claude Code Review generally available to Team and Enterprise customers on October 15. The service runs several agents on a pull request, adds a separate step to filter out false positives, and never approves or blocks a merge itself.14 About two weeks earlier, on September 28, GitHub changed Copilot's built-in review default from Lite to Balanced. Teams can now trigger reviews through the REST and GraphQL APIs, and an opt-in preview lets Copilot submit approving reviews.14 Cursor's Bugbot has been billed by usage since May. In one October hands-on comparison it ranked first on high-severity and security bugs.14 GitHub reported in May that Copilot code review had passed 60 million reviews and that more than one in five reviews on the platform involved an agent.

The editors are adding review features as well. In mid-October, Zed, Cursor, VS Code and Claude Code each added or expanded ways to accept or reject AI changes hunk by hunk. That arrived close to the Claude Code 2.1.302 release on October 17, which added client-side automatic model selection.15 Prices are also moving toward metered usage. GitHub switched all Copilot plans to usage-based AI Credits on June 1, which raised costs for heavy agent users.16 That matters here because every automated review now has a direct cost.

What engineering leaders should take from it

My reading is that the Harvard paper does not show AI coding tools are a failure. It shows they are being measured at the wrong point in the process. The authors say managers who track lines of code or commits may miss the actual return, and that capturing gains requires matching investment in the stages after coding.10 Practitioner guidance has reached the same conclusion: plan review capacity explicitly as a resource and track merge time and rework, not output volume.826

The vendors' push into review is aimed at the right problem, but the evidence that it works is still thin. The paper found AI reviewers already present in most firms without clearing the queue.23 Separate research on more than a million reviewed pull requests found that agent involvement sped up review decisions but did not improve review quality. Until a tool can take on real responsibility for approving changes, which none of the current products is designed to do, the main limit on shipping software will be how much code people can carefully review.