AI Coding Agents Write More Code, But Software Output Stays Flat

Harvard researchers Fiona Chen and James Stratton looked at data from more than 700 IT companies, covering the period from 2021 through March 2026. The sample included 300 million work events and over 700,000 employees. After companies added AI coding agents — AI tools that write code on their own — total code output rose by 30%. The number of commits, meaning saved code changes, rose by 20%. The number of finished tasks and features barely moved. Teams track these in tools like Jira, task-tracking software used by many development teams. Code review, the stage where a person checks code before it gets approved, took 49% longer on average. The share of pull requests, meaning code changes submitted for approval, that got sent back for revisions nearly doubled.

This review stage eats up the time AI agents save on writing code. A person still has to check and approve an agent's code before the work can move forward. Where companies add agents without changing the review process, the time saved on writing just shifts onto the reviewers. The amount of finished software does not grow.