AI Coding Agents Produced More Code While Human Reviews Took Longer

AI coding agents increased the amount of code developers produced without a statistically significant increase in completed software work, according to a study of more than 700 software firms reported on October 12. The time between submitting proposed changes and merging them into a codebase grew by 49 percent on average. The researchers found little evidence of increased software output or reduced employment after adoption. [4]

The important change was where the work accumulated: faster code generation pushed more work into human review rather than removing it from software production. That distinction matters to engineering managers assessing productivity, developers submitting changes and maintainers deciding what enters shared software. The study separated the activity those people could count immediately from the completed work they were trying to deliver. [4]

More code entered the queue

On October 11, The New Stack reported another gap between automated findings and completed changes: Claude had found 29,000 possible bugs in open-source software, but only 516 had been fixed. Those were possible bugs, not 29,000 confirmed defects awaiting a routine correction. The headline’s figures distinguished finding something potentially wrong from getting a fix accepted—the same distinction between an intermediate result and completed work that mattered in the firm-level research. [6]

The study reported on October 12 examined actual coding practices across more than 700 firms. After AI coding agents were introduced, total lines of code increased by 30 percent, commits increased by 20 percent and pull requests increased by 23 percent on average. Pull requests are proposed changes submitted for review before entering a codebase. All three measures recorded more activity at the production end of the process. [4]

The researchers then examined resolution rates for Issues and Epics, the work items and larger features tracked in tools such as Jira. Those rates did not change in a statistically significant way after the tools were introduced. The researchers also found no shift in the size or complexity of the tracked items that explained the result. The additional coding activity was measurable; a corresponding improvement in that completion measure was not. [4]

Review absorbed the coding gains

The discrepancy appeared in the review process. Average time from a pull request’s submission to its merger into the codebase increased by 49 percent after AI agents were introduced. The share of pull requests receiving requests for changes nearly doubled, while comments per pull request increased by 35 percent. Reviewers were having more exchanges over the proposed changes, alongside taking longer to get them accepted. [4]

Firms also spread that work across more people. The researchers found a 14 percent increase in the share of workers performing code reviews after AI agents were introduced. Their explanation was that efficiency gained during coding was absorbed by constraints later in production. Review was therefore part of the productivity calculation, rather than an administrative step that could be left out once the additional code had been counted. [4]

Automated review was already common, making this more than a case of firms overlooking another available tool. By March 2026, 80 percent of the measured firms used some form of AI code review. AI agents supplied 23.3 percent of review comments and accounted for 10.8 percent of pull requests. Humans still performed most of the work, and the researchers described AI’s contribution to relieving the review constraint as marginal so far. [4]

Teams faced different practical choices

For an engineering manager comparing an activity dashboard with a delivery tracker, the study supplied a concrete distinction. A rise in lines, commits or pull requests answered a different question from the resolution rate for tracked work. Looking at review duration, requests for revisions and completed issues alongside those activity counts offered a way to distinguish faster generation from faster delivery. The measured firms had more of the former without a statistically significant improvement in the latter. [4]

For a developer deciding how to use an assistant, Linux creator Linus Torvalds supplied a recognisable example at Open Source Summit Europe. He had written a guitar-pedal interface in C for a small screen and wanted a more polished appearance. His first AI-prompted result was poor, so he returned with the C implementation he had already written. His account described an iterative aid for work outside his strengths, grounded in an existing implementation rather than acceptance of the first generated answer. [3]

A contributor preparing an open-source submission also faced rules that differed by project. System76 banned AI-generated content from its COSMIC desktop, while NetBSD had banned AI-generated code in 2024. Debian’s developers chose to permit AI-generated contributions in August 2026. Checking the receiving project’s policy was therefore a practical step before submitting generated work. The same method could produce a permitted contribution in one project and an excluded submission in another. [2][3]

Completion and employment remained unsettled

Could automated review relieve more of the constraint, and how much checking did its own output require? The study recorded only a marginal contribution so far, despite widespread adoption. Torvalds described both sides of the problem: AI tools helped identify security vulnerabilities and improve the Linux kernel, but questionable bug reports and superficial fixes also overwhelmed maintainers. His account did not quantify the balance between useful findings and additional checking work. The number of submissions alone could not settle that question. [3][4]

What did the results establish about jobs? The researchers examined active workers in Jellyfish data and cross-referenced LinkedIn information at the firms, but could not attribute significant employment changes to AI. Separately, Kroger announced plans to cut 3,000 corporate jobs over three years as it expanded AI and other technology, according to the October 13 report. The supermarket chain employed 409,000 people. That corporate restructuring announcement and the software-firm findings concerned different workforces and different evidence; neither supplied a reason to treat announced cuts as a measured consequence of coding-agent adoption. [1][4]

Which contribution policy produced better completed software also remained open. The Register reported on October 13 that NetBSD had received a record 120 donations in the previous month, totalling $120,282 against a $50,000 target, amid increased interest in its AI-free stance. That was concrete support for a project with a clear boundary on generated code. The donations did not measure software completion or establish how its policy compared with projects permitting AI contributions. Funding, contribution rules and delivery outcomes remained distinct things to assess. [2]

What it adds up to

We think the study’s most useful finding was its separation of code production from software delivery. The extra code, commits and pull requests were real. So were the longer reviews, additional comments and wider participation in checking the work. Counting only the first stage treated an unfinished proposal as though it had already passed the people responsible for accepting it. The tracked completion results supplied an important check on that assumption. [4]

Torvalds’s experience showed why the distinction mattered without requiring a blanket verdict on AI programming. A developer could find the tool useful and enjoyable while maintainers faced more material to assess. Projects could permit generated contributions or exclude them. The central question was more specific than whether AI could write code: could the people responsible for accepting that code turn the additional submissions into completed work? Across the measured firms, the study found more coding activity but little evidence of more finished software. [2][3][4]

Compass found this story, and the sources cited here are the articles it pointed to. Compass is a service we built at Enginerds that reads 140-plus technology publications every day and keeps score of what they cover: which companies, products and technologies are drawing attention, which way each one is heading, and the dated articles behind every number. This piece began as a signal on that scoreboard, a subject the press was suddenly writing about far more than usual, and the research started from the sources Compass had already filed under it. Anyone can look up a company, a product or a technology and read its record the same way. There's a free tier at enginerds.com/compass.