AI Wrote More Code. Why Did Finished Work Barely Move?
AI coding agent(AI coding agent)
A tool that can handle longer coding tasks with limited step-by-step help.
Code review(code review)
A person checks proposed code before it joins the shared project.
Working paper(working paper)
A research report shared before formal review is finished.
What happened
A working paper by Harvard University researchers Fiona Chen and James Stratton reports a split result. AI coding agents increased the amount of code produced. They did not increase finished software by the same amount. ComplexDiscovery’s October 13 article reported the August version: code rose about 30%, finished Jira issues did not rise measurably, and pull requests took 49% longer to pass review. A pull request is a recorded request to add a code change to a shared project. Read the article.
The linked working paper now lists a November 14 version. That version reports a 44% increase in code lines after agent adoption and a 56.1% increase in time from review submission to approval. The versions should not be mixed.
The workflow behind the result
Software production has several steps. Teams choose a task, write code, review and test it, then deploy it. Code review is the human check before a change enters shared software. An AI agent can speed up drafting. It does not automatically create more reviewer time.
The paper’s model gives two possible pressures. More code can send more work to a fixed reviewer pool. More requested changes and comments can also mean that drafts need more correction. The data show the review queue growing. They do not prove that every company has the same cause.
Why it matters
The headline lesson is about measurement. Lines of code, commits, and pull requests show activity during production. A feature becomes finished only after review, testing, and deployment. If one step speeds up, another step can limit total output.
This means firms should measure approved changes, reviewer hours, rework, and delivered work. A faster drafting stage can shift the constraint instead of removing it. The study is useful because it connects early activity with later progress.
What the study measured
The current paper uses 445 million work events from 720,542 workers at 655 firms. The period runs from January 2021 through August 2026. The data came from Jellyfish, a commercial engineering analytics service. The firms consented to research use. The data connected coding activity, Jira issues, calendars, and use of GitHub Copilot, Cursor, and Claude Code.
After agents arrived, code lines rose 44%, commits rose 36%, and pull requests rose 35.7%. Jira issues rose 13.7%, while larger epics rose 18%. The paper calls the issue result borderline significant. Review time rose 56.1%. Review backlog time rose 59.9%. Active review time rose 58.5%. Comments per change rose 38.6%.
AI review tools appeared in 79.1% of firms. Humans still wrote most reviews and comments. The story also drew strong attention on Hacker News. That attention measures community interest, not research accuracy.
What remains uncertain
This is a job-market paper, not a completed peer-reviewed result. The data are proprietary and come from consenting Jellyfish clients. The authors say the data partner checked only for disclosure of proprietary information, not scientific content.
The design cannot show every cause of adoption or every kind of software quality. It also cannot settle long-run employment effects. The paper finds no detectable employment effect in its observed period. That is not proof of no future effect. The November revision also shows why dates and versions matter.
What to watch next
The useful test is whether teams publish downstream measures. Watch approved work, review hours, change requests, defects, and time to production. Watch whether more people can review safely, or whether the same senior reviewers receive more work.
AI has made one part of coding faster. The unresolved question is whether organizations can make the whole path to shipped software faster.
More AI Code Did Not Mean More Finished Work
📰 Full story: AI Wrote More Code. Why Did Finished Work Barely Move?
AI can write code quickly. People still need time to check it.
AI coding agent(AI coding agent)
An AI tool that can handle longer coding tasks.
Pull request(pull request)
A request to add a code change to a shared project.
Jira(Jira)
A tool that tracks work and marks tasks complete.
💡 The gist
- AI agents helped developers write more code.
- Finished work did not grow at the same pace.
- Human checking became the slow step.
What did researchers study?
Harvard University, a research university, has researchers named Fiona Chen and James Stratton. Their working paper studied AI coding agents.
An AI coding agent can handle longer coding tasks. ComplexDiscovery described an earlier paper version. It reported about 30% more code. Finished Jira issues did not rise clearly. Pull requests took 49% longer to pass review. Jira is a tool for tracking work. A pull request asks to add a code change to a shared project. Source article
Why did this happen?
Software work has several steps. Someone writes code. Other people review it. They test it. Then the team releases it.
Code review means people check proposed changes. AI can make writing faster. It cannot automatically give reviewers more hours. More code can therefore create more reading and checking. If the review team stays the same, waiting can grow.
Later, the paper was updated on November 14. Its newer numbers are different. Readers should check the date of the version they use. Updated paper
What is still uncertain?
The study uses records from many companies. It does not describe every company. It is also a working paper, not a completed peer-reviewed study. Its data came from Jellyfish, a commercial analytics service, and consenting clients.
The paper found no detectable employment effect during its study period. That does not prove what will happen later. It also does not answer every question about software quality.
The story received strong attention on Hacker News, a technology news site. Attention shows interest, not truth. Hacker News thread
What should we watch?
Do not count only lines of code. Watch finished features, review time, requested changes, and time to release. The important question is simple. Can AI speed up the whole process, not only writing?
💬 More AI-written code can mean more checking
HN commenters disagreed about whether AI moves the bottleneck to code review. Some described personal workplace experiences, while others pointed to requirements, design, testing, or CI/CD; none of this proves one answer for every company.
- One commenter reported that AI made each development and review task faster, but also increased requests, code, and pull requests, leaving more work for later stages. That is a personal report, not a universal result.
- Others said the harder step may be deciding what to build, designing the system, testing it, or running the automated delivery process, rather than reviewing code alone.
- People who know a company’s context and the whole system are still important. Small changes can look fine while solving the wrong problem. LLMs can help with routine checks and diagrams, but not replace human judgment.
- Another personal report described developers approving pull requests made by their own agents. No outage had happened yet in that report, but the commenter worried about major outages.
- Better tests and safety rules could reduce manual review, but they take time to build. Having one AI check another and auto-approve is not automatically safe.
initial digest at 232 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
Why Did the Code Wait?
📰 Full story: AI Wrote More Code. Why Did Finished Work Barely Move?
A computer helper writes instructions. A person still checks them.
AI coding agent(AI coding agent)
A computer helper that writes program instructions.
Hacker News(Hacker News)
A technology news site where people discuss stories.
What happened?
An AI coding agent is a computer helper that writes program instructions. Fiona Chen and James Stratton are Harvard University researchers. They studied companies using these helpers.
The helpers wrote faster. People still checked the code. Checking took longer. Finished work did not grow as quickly.
Why?
Code is a set of instructions for a computer. AI can write many instructions. A person must read them before use. More instructions can mean more reading. People still have only so much time.
Hacker News is a technology news site. Many people noticed this story there. That shows attention, not truth. Story thread
The research report is still being worked on. It does not tell us what happens everywhere. Research report
Next, people will watch more than writing speed. They will watch how quickly checked code becomes useful software.
💬 Who checks the code made by AI?
AI can make code quickly. But if it makes much more code, people may have much more to check.
- One person said their workplace became faster with AI, but also received more requests and pull requests. That is one person’s experience, not proof for every workplace.
- Other commenters said the hard parts can be choosing what to build, understanding the whole system, and testing it. A tiny change can look fine while the big plan is wrong.
- AI can help check code, but people still need to use their judgment. Some developers reportedly approved their own AI-made changes, which could hide mistakes.
- Strong tests and safety rules may help, but they take time. An AI checking another AI is not a guarantee that the code is safe.
initial digest at 232 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Does faster AI coding make code review the new bottleneck?
HN commenters split over the article’s claim that coding agents move the bottleneck from writing code to reviewing it. The discussion is mostly based on workplace reports and opinions, not independent validation of the research or proof that the pattern applies to every company.
initial digest at 232 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.