Study says AI coding agents create more code but not more software

In short: A large study found that AI coding agents help teams write more code, but they also slow down human review, so companies do not deliver more finished software.

What's going on

Harvard researchers Fiona Chen and James Stratton studied software work across more than 700 companies from 2021 through March 2026. They used data from Jellyfish, a tool that tracks engineering activity, covering about 700,000 employees and 300 million work events like code updates and review requests.

The study compared what changed after companies started using two types of AI tools. One type is an AI coding assistant, which suggests or completes code while a person is writing. The other type is an AI coding agent, which can write and submit code on its own based on instructions, a bit like an intern who can draft quickly but still needs checking.

After AI coding agents were introduced, companies produced more raw code. The researchers report a 30% increase in lines of code, a 20% rise in commits (saved code changes), and a 23% increase in pull requests (requests to add code to the main project).

But the number of finished work items did not meaningfully change. The study looked at issues and epics in Jira, which are tracked tasks and larger features, and found no statistically significant improvement.

The main reason was review time. The average time from a pull request being submitted to being accepted grew by 49%. Pull requests were much more likely to need fixes, and comments per pull request rose by 35%.

What to watch

Most firms in the study had started using AI for code review by March 2026, but people still did most of the reviewing. Watch whether better review tools, or better team habits, can reduce this “more code, more checking” slowdown.

Source: Arstechnica