AI Coding Agents Add 30% More Code but Not More Software, Harvard Study Finds

Anyone who has watched an AI coding agent work knows it can produce a lot of code quickly. A new study suggests that is close to the whole story at the firm level. Harvard researchers Fiona Chen and James Stratton find that AI coding agents increase the amount of code a company produces, but that the gain is "absorbed by downstream constraints in the production process" — chiefly human code review — with "little evidence that firms increase software output or reduce employment."

The paper leans on unusually close-to-the-ground data. It uses aggregated engineering analytics from Jellyfish, covering roughly 300 million individual "work events" — commits, pull requests and issue-management activity — across more than 700,000 employees at over 700 software development firms, from 2021 through March 2026. The researchers separate two kinds of tools: AI coding assistants, which complete code primarily written by humans, and AI coding agents, which write and submit code autonomously from prompts. Using directly measured AI usage, GitHub activity and a difference-in-differences regression around each firm's adoption, they compare what changed before and after.

On raw output, the effect is stark. Introducing AI coding agents leads to a 30 percent increase in total lines of code, a 20 percent rise in commits and a 23 percent increase in pull requests, on average.

That extra code does not translate into more shipped software. The resolution rate for Jira Issues and Epics — the tracked tasks and larger features that represent delivered work — showed no statistically significant change after AI tools arrived, and the researchers found no shift in the size or complexity of those items that would explain the gap.

The reason sits in review. After AI agents are introduced, the average time between a pull request being submitted and being merged balloons 49 percent. The share of pull requests with changes requested nearly doubles, and the number of comments per pull request rises 35 percent. In response, the share of workers performing code reviews climbs 14 percent.

Humans still do most of that work. By the March 2026 cutoff, 95 percent of the studied firms had implemented coding agents and 80 percent used some form of AI code review — yet AI accounted for only 23.3 percent of review comments and just 10.8 percent of pull requests. AI's own contribution to speeding up review, the authors found, remains marginal so far.

On jobs, the study is notably quiet. Cross-referencing Jellyfish data with LinkedIn, the researchers say they cannot attribute significant employment changes to AI across total active workers at these firms.

The finding lands on a heavily marketed premise. Coding assistants are among the most widely deployed applications of large language models, and vendors pitch them as direct accelerators of software output; enterprises are making hiring and tooling decisions on that basis. Chen and Stratton's data suggests the gain shows up as more code, not more software — a classic bottleneck dynamic, where speeding up one stage of production does nothing for throughput if another stage constrains it, and the slowest link, human verification, sets the pace. The authors add a caveat: agent capabilities have improved since the March cutoff, and firms are still learning where to deploy them, so the balance between generation and review may shift.