AI coding agents are generating more code, according to a new Harvard working paper.
The paper, by Harvard researchers Fiona Chen and James Stratton, uses aggregated Jellyfish engineering data covering roughly 300 million work events across 718 firms from January 2021 through March 2026. The current version is dated Aug. 4, 2026. The authors studied both AI coding assistants, which offer suggestions while a person writes code, and newer agentic tools that can semi-autonomously complete longer coding tasks. The study uses a staggered difference-in-differences design, comparing firms before and after adoption while using different adoption timing across organizations.
The study found that AI agents increased lines of code by 30%, commits by 20%, and pull requests by 23%. But Jira issues and epics, used in the paper as firm-level measures of deployed tasks, projects, or features, did not show statistically significant gains after adoption.
The code review queue absorbed much of the gain
Chen and Stratton point to code review as the main constraint. After firms adopted AI agents, the average time from pull request submission to merge rose by 3.45 days, a 49% increase from the baseline. The share of pull requests with change requests increased from 13% by 12 percentage points, nearly doubling. Review comments per pull request rose 35%.
The shift also changed how people were allocated. The share of workers doing code review increased 14%, while the share of workers writing code did not significantly fall. In practical terms, the firms were not simply replacing engineering effort with agent output. They were moving more human attention into checking that output.
The study’s time-allocation measures help explain why a boost in coding alone may stall. The authors estimated from Jellyfish data that engineers spent 42% of their time writing code and validated the broader point with a survey of 100 Prolific engineers, who reported spending about 30% of their time coding. The rest of the software production system, including planning, review, testing, coordination, and release, still has to keep pace.
AI review tools are present, but humans still own the gate
The paper also tested whether AI review tools had already loosened the constraint. By March 2026, nearly 80% of firms in the sample had used AI tools in the code review process. Yet only 23.3% of all review comments were generated by AI, and only 10.8% of pull requests received at least one AI-generated comment.
That distinction matters. AI tools may be widely available in review workflows, but the paper’s data still shows human reviewers doing most of the verification. The authors argue that review can remain a bottleneck when AI increases the amount of draft code entering the pipeline or changes the amount of scrutiny each submission requires.
A broader pattern is emerging
The Harvard firm-level study is not the only evidence separating code production from software delivery. A September 2026 revision of an NBER working paper by Mert Demirer, Leon Musolff, and Liyuan Yang used data from more than 500,000 GitHub developers combined with AI usage telemetry. The authors reported cumulative increases in coding activity of about 30% with autocomplete tools, 180% after adding interactive coding agents, and 240% after adding autonomous coding agents.
Those gains were much smaller later in the production chain. The same authors reported that the 240% cumulative increase in coding activity fell to about 80% for the number of projects and about 30% for actual releases. Their broader analysis also found that large coding gains translated only partially into software that was ultimately shipped and used.
The two studies are not identical. The Harvard paper looks at firm adoption using proprietary Jellyfish data and uses Jira issues and epics as output measures. The NBER paper tracks developer and marketplace activity. But both point in the same direction: activity metrics near the code-writing stage can rise much faster than the measures closer to delivered software or user adoption.
DORA links AI adoption to higher throughput and weaker stability
Drawing on survey responses from nearly 5,000 technology professionals, Google Cloud’s 2025 DORA report described AI as an amplifier of existing strengths and weaknesses. Google Cloud said AI adoption was associated with higher delivery throughput and product performance in 2025, a reversal from its 2024 findings, but continued to have a negative relationship with software delivery stability.
What technology teams should measure next
AI adoption should be measured across the whole delivery path, not only at the point where code is produced. Agent seat counts, generated lines, accepted completions, commits, and pull requests may show activity. They do not prove that delivery has improved.
For teams using or evaluating coding agents, the next set of measurements should sit closer to the actual constraint:
- time from pull request opened to merged;
- share of pull requests that need changes;
- review comments per pull request;
- issues, epics, or roadmap items completed;
- change failure rate, incident volume, rollback rate, and rework;
- customer usage or adoption of shipped software.
Those measures are less exciting than a chart showing more code. They are also closer to the business question: whether AI is increasing useful output or increasing the amount of work that must be reviewed before useful output appears.
The jobs finding is narrower than the headlines may suggest
The Harvard paper also found no significant employment change that could be attributed to AI adoption in the sample. The authors measured total employment using LinkedIn profiles connected to firms and engineering employment through active workers in Jellyfish data. For AI agents, they reported a near-zero effect on overall employment and a noisy zero effect on engineering employment.
That should not be read as proof that AI will never affect software jobs. The authors explicitly framed the result as a short-run estimate and noted that employment may respond slowly. But it does weaken the idea that, in this firm sample through March 2026, agent adoption had already translated directly into fewer software workers.
I can suggest related Tech Help Canada articles based on the topic you’re reading now.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
