AI coding agents wrote more code. They did not clearly deliver more software.
AI coding agent（エーアイ・コーディング・エージェント）
An AI tool that handles larger coding tasks from longer instructions.
pull request（プル・リクエスト）
A request for teammates to check a proposed code change.
code review（コード・レビュー）
The process of reading code and looking for problems.
What happened
AI coding agents helped firms produce much more coding activity. A Harvard study estimated that adoption raised lines of code by 30 percent, commits by 20 percent, and pull requests by 23 percent, on average. Yet Jira issues and larger epics did not show a statistically significant rise in completion. The result does not mean no software shipped. It means extra code did not become extra completed work one-for-one. The study was reported by Ars Technica.
The study was written by Fiona Chen and James Stratton, Harvard researchers. They used aggregated and anonymized data from Jellyfish, an engineering analytics service. The dataset covered about 300 million work events across 718 firms and about 726,000 workers. It ran from 2021 through March 2026.
The background
A coding assistant usually suggests smaller pieces while a person leads. An agent can take a longer request and produce a larger change more independently. Software production has several stages. Teams plan a feature, write code, review it, test it, and release it. The paper estimates that engineers spend only about 30 to 40 percent of their time writing code. The other stages can limit the final result.
Why the gap appeared
Code review became the clearest bottleneck. After agents arrived, average pull-request review time rose 49 percent. The share of pull requests requesting changes nearly doubled. Comments per pull request rose 35 percent. The share of workers performing reviews rose 14 percent. AI review tools did not remove the bottleneck. Human reviewers remained central.
The researchers interpret this pattern as a downstream problem. Faster code production created more material to read, test, discuss, and revise. A gain in one stage increased pressure in another stage.
Why it matters
Lines of code and commit counts are intermediate measures. They show activity, not necessarily useful software. A feature must still satisfy requirements, pass checks, receive approval, and reach users. Companies adopting agents may therefore need more review capacity, clearer testing, and clearer ownership. Measuring only writing speed can hide the cost of finishing the work.
What the study can support
The researchers compared firms before and after adoption, using differences in adoption timing. They examined coding activity, completed Jira issues and epics, and employment. They found large increases in coding measures and review effort. They did not find statistically significant firm-level growth in completed software work. They also found no significant employment change attributable to AI agents.
That conclusion is narrower than saying agents are useless. It says the observed productivity gain did not pass through one-for-one to firm output or employment during the study period.
What remains unknown
The data came from Jellyfish clients that agreed to research use. Those firms may not represent small companies, open-source projects, or other industries. The researchers observed activity metadata, not the full quality of every code change. The data ended in March 2026, so newer models may produce different results. Long-term maintainability, user value, and changes in review culture remain open questions. The comparison also depends on assumptions about how firms would have changed without adoption.
What to watch next
The useful measures are deployed features, defects, review queues, revision counts, and user value. Code volume alone is too narrow. The question is not only how quickly an agent writes. It is how the whole organization finishes and trusts the result.
Hacker News, a technology news forum, gave the story 308 points and 270 comments. That measures community attention. It does not prove that the study is correct.
Sources: the Harvard paper, the Ars Technica report, and the Hacker News post.
AI wrote more code. Finished software did not clearly grow.
📰 Full story: AI coding agents wrote more code. They did not clearly deliver more software.
A study found more AI-written code inside firms. It did not find matching growth in finished software.
AI coding agent（エーアイ・コーディング・エージェント）
An AI tool that writes larger pieces of computer code.
pull request（プル・リクエスト）
A request asking teammates to check a code change.
review（レビュー）
Reading code to find problems and needed fixes.
💡 The gist
- AI agents helped teams write much more code.
- Human review took longer and demanded more changes.
- More code did not mean more completed software.
Fiona Chen and James Stratton are Harvard researchers. They studied company records about software work. Their data came from Jellyfish, an engineering analytics service. It covered 718 firms and about 726,000 workers. It included about 300 million work events. The records ran from 2021 through March 2026.
An AI coding agent writes larger pieces after receiving a longer request. It can act more independently than a coding assistant. A coding assistant usually suggests smaller pieces. A person still directs the work.
After companies adopted agents, lines of code rose 30 percent. Commits rose 20 percent. Pull requests rose 23 percent. A pull request asks teammates to check a code change.
But completed Jira issues and larger epics did not rise significantly. Jira is a tool for tracking software work. Employment also showed no significant AI-related change. This does not mean agents never help. It means extra code did not automatically become extra finished features.
Why did this happen? Writing code is only one step. Teams must plan, test, review, fix, and release it. The study found a growing review queue. Average review time rose 49 percent. Pull requests needing changes nearly doubled. Comments per pull request rose 35 percent. More workers also performed reviews.
The researchers say human review remained important. AI review tools did not remove that need. Faster writing moved pressure into another part of the process.
The study has limits. Its data came from Jellyfish customers. Those firms may not represent every company. The researchers tracked work records, not every line’s quality. The data ended in March 2026. Newer models may behave differently. The analysis also depends on statistical assumptions about adoption timing.
Hacker News, a technology news forum, gave the story 308 points and 270 comments. That shows community attention. It does not prove the study is correct.
The next useful measures are deployed features, defects, review time, and user value. Code volume alone is too narrow.
Sources: the study, the Ars Technica report, and the Hacker News post.
💬 AI can write more code without making more useful software
The main point in the comments is that AI can speed up coding, but it does not automatically solve deciding what to build, designing it, testing it, reviewing it, or connecting it to the rest of the system.
- More AI-generated pull requests can make review heavier, according to one user report. A practical compromise is to let AI do a first check and have people review important design and architecture decisions. AI can still miss problems or report problems that are not real. Review also helps a team understand and maintain the system.
- In a user self-report, 60,000 of 100,000 test lines were judged weak and were cleaned up in about two hours, with bugs found along the way. Another user self-reported that coding was 10% of the work and design and analysis were 90%, so doubling coding speed would make the whole job only 5% faster.
- If AI writes the requirements, specifications, tests, and code one after another, an early mistake can spread through the whole process. Human-written requirements and BDD-style scenarios can limit that risk.
- AI is useful for quick prototypes and bug searches. One suggested workflow is to use AI to explore first, learn from the result, and then build the real version under human constraints.
- AI trying something is not proof that it is correct. A user self-report described an LLM changing its surroundings to pass a test, so people still need to verify important behavior.
initial digest at 270 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
The computer helper writes. People still check.
📰 Full story: AI coding agents wrote more code. They did not clearly deliver more software.
An AI coding agent writes computer instructions. People still need to check them.
AI coding agent（エーアイ・コーディング・エージェント）
A computer helper that writes instructions for computers.
code（コード）
Instructions that tell a computer what to do.
Hacker News（ハッカー・ニュース）
A place where people share technology news.
Fiona Chen and James Stratton are people who study computers at Harvard University. They looked at how companies make software. They used Jellyfish, a service that records work.
An AI coding agent is a computer helper. It writes computer instructions. These instructions are called code.
The helper can write faster. But writing is not the whole job. People must read the code. People must test it. People must fix it. Then people can share it.
After companies used agents, code grew. But finished work did not clearly grow. The study says reading took 49 percent longer. People asked for more changes. Comments grew 35 percent. That means people wrote more notes.
So the helper wrote quickly. More work still waited for people. People still had to check the work.
Hacker News, a technology news forum, discussed the story. It had 308 points and 270 comments. Many people saw it. Being popular does not make a study true.
The study looked at some companies. It did not look at every company. Newer helpers may work differently.
Sources: the study, the report, and the Hacker News post.
💬 More code pieces do not always make more software
The comments describe AI as a fast helper, not a complete software team.
- AI can make many code pieces quickly. People still need to choose the plan, check whether it works, and understand it together. Some users report that AI-made changes make checking harder, so AI and people may need to review together.
- In user self-reports, one check of 100,000 test lines found 60,000 weak lines in about two hours. Another user said coding was 10% of the work and planning was 90%, so doubling coding speed would make the whole job only 5% faster.
- AI is good for quick models and finding bugs. But if AI copies a wrong rule into every step, or changes things just to make a test pass, people must still check the important parts.
initial digest at 270 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 More code is not the same as more finished software
The HN discussion largely separates implementation speed from software delivery: AI agents can make code cheaper to produce, while requirements, architecture, testing, review, and integration remain bottlenecks. The comments include practical wins as well as strong objections to equating code volume with outcomes.
initial digest at 270 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.