AI coding agents raise code output. In a Harvard working paper revised in November 2026, pull requests rose 35.7% a year after adoption, while resolved issues rose 13.7%, a borderline result. More code arrives and the same reviewers have to judge it: review time rose 56.1% and time waiting for a reviewer rose 59.9%.
What are we evaluating?
Review basis. Production experience. Conclusions drawn from practice in production, with the data we may disclose.
Methodology. At ACW Soft Solutions, AI agents write the code for our standard development tasks, and we built the process around that step. In our runs an agent spends 41 minutes writing the code and 51 minutes planning and checking it before a person opens the pull request. Both figures are medians for standard tasks between 28 October and 5 November 2026.
Tested on. 5 Nov 2026.
Evidence. [MISSING: link to or reference for ACW's run data behind the 41- and 51-minute medians for standard tasks, 28 October to 5 November 2026]
More code, not more finished product
Companies that adopt AI coding agents open 35.7% more pull requests a year later, and each pull request takes 56.1% longer to review. The figures come from a Harvard working paper revised on 7 November 2026, which tracks 720,542 engineers at 655 firms. Finished work moved much less: resolved issues rose 13.7%, a result the authors describe as only borderline significant.
For a business, that is the gap between what AI coding costs and what it delivers. The budget buys more code. Whether it buys more finished product is decided at the step after the code is written. In Sourcegraph's 2026 Big Code survey, published on 16 November 2026, 54% of respondents at AI-using organisations estimated that more than 60% of AI-generated engineering work needs significant human rework before release.
At ACW Soft Solutions, AI agents write the code for our standard development tasks, and we built the process around that step. In our runs an agent spends 41 minutes writing the code and 51 minutes planning and checking it before a person opens the pull request. Both figures are medians for standard tasks between 28 October and 5 November 2026.
This article covers why review slows down, the five changes we made, and what to measure if your company pays for AI coding.
Why AI coding agents make code review slower
AI coding agents make code review slower because they raise the volume of code while the number of people able to judge it stays the same. In the Harvard data, lines of code rose 44.0% and pull requests 35.7%. Time spent waiting for a reviewer rose 59.9%.
Companies responded by moving people. Effort spent on coding fell 22.7%, and effort spent on review rose 43.4%. AI review tools did not remove the queue: 79.1% of the firms adopted them, and most reviews and comments were still written by people.
The paper has not been peer reviewed, so the exact percentages deserve care. The mechanism it describes is the one we designed our own process against.
Five changes that keep review from becoming the queue
Each change answers one reason AI-written code waits for a person.
1. Two reviews happen before a person is asked to look
When a person is the first reader of AI-written code, that person does all the checking. In our process the agent that wrote the code never reviews it, because an agent that has just written something tends to consider it correct. Separate agents review the change twice, and the whole change is then read once more. A serious finding counts only after a fresh agent confirms it.
2. A later review round checks only the fix
Review rounds multiply if each one rereads everything. Before this rule, every round found new small remarks and review went round four or five times. Now a later round fails only on an earlier finding that is still open or on a serious new problem. Review stops at two rounds, with a third only when one to three minor remarks remain.
3. Work is cut to a size a person can read
A task estimated at more than eight hours is flagged when it is generated and split before it starts. Such a task usually returns as a pull request of a thousand lines or more, and nobody can check that by reading it.
4. The pull request tells the reviewer where to start and what is missing
Every pull request opens with a summary of what was done, a reading order, the vulnerable packages the change brought in, and a list of what was not done. If the automated checks fail three times, the run stops and reports the likely cause.
5. Tasks run in parallel, so reviewing and writing overlap
A reviewer and an agent do not wait for each other. Several tasks run at the same time, each in its own sandbox and branch. While a developer reviews one pull request, agents are already working on the next tasks. A task that depends on an earlier one starts on top of that task's pull request as soon as it is open and its checks pass, without waiting for the merge. The time a developer spends reviewing is no longer idle time for the project.
What the business gets from this
Four things change for the company that pays for the work.
- Reviewer time goes to judgement. A developer reads a change that has already passed the build, the tests and two reviews, and decides whether it is right for the product.
- No idle time between tasks. Work continues in parallel while a person reviews, so the project does not stand still waiting for code to be written.
- A unit to plan with. At the median, a standard task reaches a checked pull request in about an hour and a half.
- A named owner. The developer who merges answers for the code. That responsibility does not move to the agent.
The limits stay. Agents do not make architectural decisions, tasks that depend on each other still run in order, and a vaguely described task still produces poor code.
What to measure if your company pays for AI coding
| Measure | Why it matters |
|---|---|
| Finished work: resolved issues, shipped features | Commits and lines of code count what enters the review queue, not what leaves it |
| Time from pull request to merge | It shows whether review has become the constraint |
| Size of a typical change | A change too large to read tends to be approved without being read |
| Share of AI-written changes that need rework | Rework is where the saved coding time goes |
| Who is named on each merge | Responsibility has to sit with a person |
If your team has adopted AI coding and the review queue is growing, we can go through your task flow and checks with you as part of our AI-Native Engineering service.
What are the limitations and trade-offs?
- The paper has not been peer reviewed, so the exact percentages deserve care.
- Agents do not make architectural decisions, tasks that depend on each other still run in order, and a vaguely described task still produces poor code.
Frequently asked questions
Do AI coding agents make software teams faster?
AI coding agents raise code output. In a Harvard working paper revised in November 2026, pull requests rose 35.7% a year after adoption, while resolved issues rose 13.7%, a borderline result.
Why does code review take longer with AI coding agents?
More code arrives and the same reviewers have to judge it. In the same data, review time rose 56.1% and time waiting for a reviewer rose 59.9%.
Do AI code review tools remove the bottleneck?
Not by themselves. 79.1% of firms in the Harvard sample adopted AI review tools, and most reviews were still written by people. At ACW Soft Solutions agent review comes before a developer's review and does not replace it.
How large should a task for an AI coding agent be?
At ACW Soft Solutions a task estimated at more than eight hours is split before it starts, because it usually returns as a pull request of a thousand lines or more.
Who is responsible for code written by an AI agent?
The developer who merges it. In ACW's process an agent opens the pull request and a person decides whether it is merged.
Does a developer's review hold up the next task?
Not at ACW Soft Solutions. Tasks run in parallel, each in its own sandbox and branch, and a dependent task starts on top of the earlier pull request without waiting for the merge.
About the author. Oleksandr Veklych, Co-founder & CTO, ACW Soft Solutions. Oleksandr Veklych is the CTO and co-founder of ACW Soft Solutions, where he leads technical strategy and delivery for custom software: SaaS platforms, marketplaces, internal business systems, mobile apps and AI-based solutions. His focus is turning growing complexity into clear architecture, reliable delivery and scalable technical decisions, so teams can move faster without rebuilding everything. He holds a bachelor’s degree in cybersecurity from Taras Shevchenko National University of Kyiv. View profile

