Future of work
AI coding agents produced more code, not more finished software
- Published
- Source
- Ars Technica
Summary
Harvard researchers analyzed engineering analytics from Jellyfish covering more than 700 software firms from 2021 through March 2026. After AI coding agents were introduced, code output grew, but issue and feature completion rates showed no significant gain, while average review time per pull request rose 49 percent and the share of pull requests with changes requested nearly doubled. The researchers found no significant employment changes they could attribute to AI. This summary is based on Ars Technica's report; the paper itself is available as a PDF.
Why it matters for our work
The study shows that human review time grows together with machine-generated output, making the generation-review balance central to any productivity claim. AI adoption should be measured by finished work rather than raw code volume.
Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.