AI Coding Agents Generate More Code, but Reviews Can Slow Software Delivery
AI coding agents can speed up code creation without increasing software delivery at the same rate. A study cited by Ars Technica found that, after teams adopted these tools, the average time between submitting a proposed change and merging it into a project increased by 49%. The finding shows that writing code faster does not, on its own, address every stage of development.
The main concern is code review. Before a change becomes part of a project, someone needs to understand what was done, check whether the solution meets the requirements, and assess its effects on the system. When that stage requires more revisions and discussion, some of the gains from faster code generation can disappear.
Producing more code is not the same as delivering more software that is ready to use. That distinction helps explain the study findings and puts expectations for AI-driven productivity into perspective.
What the Study Reveals About AI-Assisted Coding
The research by Chen and Stratton, cited by Ars Technica, identified workflow changes after the introduction of coding agents. Along with longer review timelines, the authors observed more requests for revisions and more comments on proposed changes.
The main reported findings were:
- An average increase of 49% in the time between submitting a pull request and merging it into the project codebase.
- A nearly twofold increase in the share of pull requests with changes requested after teams adopted the agents.
- A 35% increase in the number of comments per pull request.
- A 14% increase in the share of professionals performing code reviews.
A pull request is a proposed change submitted for review before it is incorporated into a project. It allows other team members to read the code, ask questions, and request revisions. Its progress can therefore help show how work moves from creation to approval.
These numbers point to greater pressure on the validation stage. However, they do not mean that every project experienced the same outcome or that every AI-generated change had problems. They reflect the analyzed dataset and should be interpreted within that context.
Why the 49% Increase Requires a Careful Reading
The time measured in the study covers the interval between submitting a pull request and merging it into the project. That is different from measuring only the minutes someone spends reading code.
This interval can include waiting for an available reviewer, exchanging comments, running tests, and completing additional rounds of revisions. In other words, reading the code does not have to become 49% slower for the overall process to take 49% longer.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
The distinction matters because it helps identify the potential bottleneck. A team can generate changes faster and still take longer to merge them. If review capacity does not keep pace with the volume of proposed changes, the review queue can grow.
This is one possible explanation for the problem, not a single proven cause across every case in the study. The excerpt cited by Ars Technica shows an increase in turnaround time and interactions, but it does not detail how much each factor contributed to the result.
More Comments and Revision Requests Add to Validation Work
The 35% increase in comments per pull request indicates more interactions during review. Meanwhile, the nearly doubled share of proposals with changes requested shows that a larger portion needed revisions before moving forward.
A comment, however, does not necessarily indicate an error. It may ask for an explanation, suggest a simpler approach, or clarify a product rule. Likewise, a change request can involve anything from a minor adjustment to a substantial reworking of the solution.
Even with those differences, each round of discussion can add work to the process. The author needs to understand the feedback, modify the code, and submit a new version. The reviewer then needs to check whether the issue has been addressed.
A Solution That Looks Correct Still Needs Context
In practice, reviewing software involves more than checking whether the code passes a simple test. Reviewers need to assess whether it follows business rules, preserves expected behavior, and fits the project structure.
For example, a change to a discount calculation might produce the correct amount in the most common scenario while overlooking a rule specific to a particular plan. This example illustrates why context matters; it is not a case reported in the research.
AI agents can help create functions, edit files, and suggest tests. Still, the speed of those actions does not eliminate the need to verify that a change solves the right problem.
Human Participation in Code Review Increased
The researchers observed a 14% increase in the share of professionals involved in reviews after the agents were introduced. The finding suggests that tool adoption was accompanied by a shift in how work was distributed, with more people participating in this stage.
This should not be confused with a 14% increase in hiring. The figure refers to the share of workers performing reviews, not an equivalent expansion in team size.
After analyzing the total number of active professionals in the Jellyfish data and cross-referencing that information with company data from LinkedIn, the authors also reported that they could not attribute significant employment changes to AI.
This result does not prove that artificial intelligence will never affect software development jobs. It simply defines what the researchers were able to identify in the analysis presented. Within the scope described, there was no basis for attributing a significant change in that metric to AI.
AI Also Reviews Code, but Its Role Remains Limited
One possible response to the growing workload would be to use artificial intelligence itself to review changes. The study indicates that this practice was already present at a large share of companies, but its contribution to the observed metrics remained limited.
By March 2026, 80% of the companies analyzed were using some form of AI code review. Despite that adoption, agents accounted for 23.3% of all review comments and 10.8% of all pull requests, according to the figures cited by Ars Technica.
These percentages measure different things and should not be added together. They are not a direct measure of hours saved, either. Still, they help explain why having the tool in place at many companies does not mean it already handles most of the work.
Adopting AI-assisted review and automating the entire review process are very different things. In the scenario reported, people remained responsible for most of that work.
How to Evaluate Productivity Beyond Code Volume
The findings reinforce an important distinction for technology teams: output metrics alone do not reveal the value delivered. Counting lines of code or opened pull requests shows activity, but it does not tell you whether the changes were approved, released, and well received by users.
A more complete assessment can consider:
- Time to merge: how long it takes for a change to be incorporated into the project.
- Revision rounds: how many times a proposal needs to go back for adjustments.
- Review effort: how much human work validation requires.
- Quality after release: whether defects emerge or the solution needs to be reworked.
- Product outcome: whether the change addresses the need that prompted it.
These criteria are practical ways to assess the development workflow, not additional findings attributed to the study. They help prevent an improvement in one area, such as writing code faster, from being mistaken for progress across the entire process.
What Can Make Changes Easier to Review
Proposals focused on a clear objective tend to be easier to understand. When a change combines many unrelated tasks, the reviewer needs to follow more decisions at once. Breaking the work into coherent pieces can make that assessment simpler.
Context also makes a difference. A description that explains the problem, summarizes the solution, and lists the tests performed helps the team compare the implementation with the original request. This applies to both human-written code and AI-assisted changes.
Automated tests and coding-standard checks can complement the review. However, they do not replace every decision about behavior, maintainability, and product fit. These practices may reduce friction, but the excerpt presented does not demonstrate that they eliminate the observed increase in turnaround time.
Companies Are Still Learning How to Use Coding Agents
According to Ars Technica, 95% of the companies in the study had already deployed coding agents in the scenario described. At the same time, the article notes that many could still have been learning when and how to use these tools most effectively.
The article also points out that agents received updates after March 2026, the data cutoff date. That limits any attempt to treat the findings as a definitive description of all tools and future versions.
With more experience, teams may identify tasks where automation offers a better balance between creation and review. That is a possibility, not a gain guaranteed by the reported numbers.
The central conclusion remains: the benefit of AI-assisted coding depends on the full path from generating a change to putting it into use. In the cited study, increased review work reduced the advantage of faster code creation. The challenge is not simply to produce more code, but to turn that output into reliable software without shifting all the effort to the next stage.
