A Harvard working paper covering 718 companies finds that AI coding agents increase how much code teams write, yet delivered output has barely changed. Time spent on code review rose by 49%.
Many teams adopted AI coding agents hoping to ship faster. According to the Harvard working paper, code volume grew by roughly 30%, but what teams actually delivered stayed almost flat. The slowdown appears to have shifted downstream to review.
The article behind this summary points to three causes. Generating code is cheap, while checking it takes far longer. Agent-written code often looks convincing and can hide subtle problems in places reviewers tend to skim. And review capacity is limited: a senior engineer who knows the domain cannot be scaled up automatically.
The author lists several practices from teams coping well. Ask the agent to produce small, self-contained diffs, since a short pull request with a clear summary is far more likely to get a real review than a huge one. Have the agent review its own diff before a human does, which can catch many obvious issues and shorten the human pass. Require the agent to write pull request descriptions that explain why an approach was chosen. And track time-to-merge rather than just lines generated.
The author argues that the finding is more a management problem than a tooling one. Organizations bought faster generation but found that verification was the real constraint. Until review gets comparable engineering attention, the extra 30% of code is largely unshipped inventory.
Source: Dev.to
