AI coding agents raise code output 30% but not shipped features

A new look at how AI coding tools are used in real teams finds a familiar pattern: more code gets written, but not more usable software ships. The study, led by Harvard University researchers Fiona Chen and James Stratton, shows that while AI coding agents can speed up writing code, they don’t automatically translate into more completed software. In plain terms: more code doesn’t necessarily mean faster or better products for users.

The researchers analyzed data from Jellyfish, a platform that tracks how engineering teams work. The data set spans 2021 through March 2026 and covers hundreds of firms, including millions of work events like commits and pull requests from hundreds of thousands of engineers. They compared teams that started using AI coding assistants (which suggest and complete human-authored code) and AI coding agents (which can generate and submit code autonomously) against teams that did not use these tools yet.

What they found: after AI agents are introduced, total lines of code rise by about 30 percent, commits increase by around 20 percent, and pull requests grow by roughly 23 percent. But the measure of actual software output, finished features and deliverables tracked in project management systems, did not show a meaningful improvement. In other words, more code does not automatically mean more features reaching users.

The authors emphasize a key constraint: the code review and integration process becomes a bottleneck that absorbs much of the potential gain from AI tooling. ReviewActivity and issue resolution did not shift in a statistically significant way after AI tools arrived, suggesting teams still need to complete the same heavy lifting before new code becomes part of a product.