What changed
A Harvard study found that AI coding agents sharply increase coding activity without producing more tracked software output. After firms adopted the agents, lines of code rose 30%, commits 20% and pull requests 23%, while Issue and Epic resolution rates did not change significantly.
The study, by Fiona Chen and James Stratton, analyzed 300 million work events involving more than 700,000 employees at over 700 software firms from 2021 through March 2026.
Why it matters
The constraint moved downstream. Pull-request review took 49% longer after AI agents arrived, and requests for changes nearly doubled. More code entered the system, but humans had to spend more time checking, revising and approving it.
That is the practical warning for teams buying these tools: faster generation can simply create a larger inspection queue. The useful measure is not how much code an agent produces. It is how much finished software survives the review process and reaches completion.
The study’s firm-level measure may not capture every form of software value, and its excerpt does not provide sector-level results. Still, the evidence points to a clear near-term tradeoff: unless review capacity grows with agent output, the productivity gain can be absorbed before it becomes a shipped feature.
Comments
No comments yet.