{"slug": "ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard", "title": "AI Coding Agents Write 30% More Code – and Resolve Zero Additional Issues, Harvard Study Finds", "summary": "A Harvard working paper by researchers Fiona Chen and James Stratton, analyzing Jellyfish engineering analytics data from 718 firms and over 725,000 workers between 2021 and 2026, found that AI coding agents such as Claude Code, Cursor and Devin raised code production 30% — nearly 1,500 additional lines per worker-month — with no statistically significant increase in resolved Jira issues or completed epics. Commits rose 20% and pull requests 23% per month, but review times jumped 49% (an average of 3.45 days), the share of pull requests requiring revision nearly doubled, and comments per request rose 35%. The paper concludes that productivity gains do not fully pass through to software output, with 14% more workers pulled into code review even as adoption climbed from near zero in late 2024 to over 95% by early 2026.", "body_md": "Enterprise technology leaders are currently caught in a measurement trap. When evaluating the impact of AI coding agents, the industry has largely focused on the volume of code produced. It is an easy metric to track, but it is increasingly clear that it is the wrong one. By prioritizing output speed, firms are ignoring the reality of how software actually moves from an idea to a finished product.\n\nA new [Harvard working paper](https://fion.ac/jellyfish.pdf), *Artificial Intelligence in the Firm: Bottlenecks in Software Production*, provides the first large-scale empirical evidence that this focus on volume is masking a significant operational hurdle. Researchers Fiona Chen and James Stratton analyzed data from the [Jellyfish](https://www.jellyfish.co) engineering analytics platform, covering 718 firms and over 725,000 workers between 2021 and 2026. The findings are stark: while AI agents like Claude Code, Cursor, and Devin have increased code production by 30%-adding nearly 1,500 lines per worker-month-there has been no statistically significant increase in resolved Jira issues or completed epics.\n\nThe data reveals a disconnect between activity and outcome. While workers are pushing 20% more commits and 23% more pull requests per month, the business value-measured by resolved issues-remains flat. As one junior engineer interviewed for the study noted, “People could make a huge number of commits quickly, but code review was still the bottleneck.”\n\nThe mechanism behind this is the code review process. When developers use AI to generate code faster, they are essentially pushing more work into the review queue. The data shows that review times have jumped by 49%, increasing by an average of 3.45 days. Furthermore, the share of pull requests requiring revision has nearly doubled, and the number of comments per request has risen by 35%. As one senior engineer noted in the study, “We were able to write code faster in some instances, but there is only one bottleneck at a time. If you speed up coding, the bottleneck shifts somewhere else.”\n\nThis phenomenon highlights a critical flaw in how we track engineering health. We are measuring the speed of the engine while ignoring the fact that the transmission is stuck. By focusing on the wrong metrics, companies are failing to see that their investment in AI is not necessarily creating more software, but rather creating more work for the people responsible for reviewing it. The authors of the paper put it plainly: “Both technologies increase coding productivity. However, productivity gains do not fully pass through to changes in software output or employment.”\n\nThe human cost of this shift is becoming increasingly apparent. The study found that 14% more workers are now being pulled into the code review process, placing a heavy burden on senior engineers who must vet the influx of AI-generated updates. This mirrors the phenomenon previously identified in our coverage of [Cognition’s Devin](https://forkast.news/cognitions-devin-writes-89-of-its-own-companys-code-the-real-story-is-whats-left-for-humans/), where the high volume of AI-written code creates a significant “babysitting tax.” Even with advanced AI tools, human oversight remains an absolute requirement, and that oversight is now consuming more time than ever. As one senior engineer emphasized, “Even with AI tools, you absolutely still need humans involved in the review process.”\n\nThe pattern is clear: firms are rushing to adopt AI agents-with adoption rates climbing from near zero in late 2024 to over 95% by early 2026-without adjusting their internal workflows. Despite 80% of firms using AI-assisted review tools, only about 23% of review comments are actually AI-generated. This suggests that the heavy lifting of quality control is still falling squarely on human shoulders, even as the volume of code to review continues to climb.\n\nFor engineering leaders, the implication is that the current AI adoption narrative needs a correction. As the authors note, “Realizing the benefits of AI adoption requires complementary investments in downstream review capacity, rather than simply increasing code output.” This might mean better tooling for automated testing, more efficient review workflows, or even rethinking how teams are structured to handle the increased volume of incoming code. If you speed up one part of a complex system without addressing the constraints of the next, you are not increasing productivity; you are just moving the pile of work.", "url": "https://wpnews.pro/news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard", "canonical_source": "https://forkast.news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard-study-finds/", "published_at": "2026-10-11 09:32:20+00:00", "updated_at": "2026-10-11 09:52:18.104430+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-products", "developer-tools"], "entities": ["Harvard", "Fiona Chen", "James Stratton", "Jellyfish", "Claude Code", "Cursor", "Devin", "Jira"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard", "markdown": "https://wpnews.pro/news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard.md", "text": "https://wpnews.pro/news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard.txt", "jsonld": "https://wpnews.pro/news/ai-coding-agents-write-30-more-code-and-resolve-zero-additional-issues-harvard.jsonld"}}