Harvard researchers Fiona Chen and James Stratton analyzed granular engineering analytics from Jellyfish covering 300 million work events across more than 700 firms (2021-March 2026) to measure the impact of AI coding assistants and autonomous coding agents. They find that introducing agentic tools spikes production metrics - about 30% more lines of code, 20% more commits, and 23% more pull requests - yet yields no statistically significant change in higher‑level software output such as resolved issues or completed epics. The researchers used difference‑in‑differences regressions tied to firms’ staggered AI adoption to isolate these effects.
The disconnect is explained by a downstream code‑review bottleneck: average time from pull request submission to merge rises roughly 49%, the share of PRs requiring changes nearly doubles, and comments per PR increase ~35%. Firms respond by reallocating reviewers (a ~14% rise in workers doing reviews), but total employment shows no meaningful change. Although most firms had some AI review tools by March 2026, AI accounted for only ~23% of review comments and handled ~10.8% of PRs, leaving humans doing most of the verification. The net result is that faster code generation is largely absorbed by greater review effort, leaving uncertain near‑term productivity or employment gains until review processes or agent quality improve.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.