Thursday, November 19, 2026 Updated Briefings Saved 中文
AI Daily Radar

AI and the frontier tech around it, explained every morning

Demonstrated

Harvard researchers link AI coding agents to a code-review bottleneck

Photo: Pexels / Pixabay

Harvard researchers have reportedly concluded that AI coding agents have moved the main developer bottleneck away from writing code and toward reviewing it, according to secondary coverage of the work. The account gives no details of the underlying study — its methodology, sample size, which models or repositories it examined, or whether it has been peer reviewed — so the strength of the finding cannot be judged from what is available. Practitioner discussion on Hacker News broadly matches the description: several commenters say agents now generate more merge-request code than before, and that reviewing it, including having agents review each other’s output, has become the bulk of their working day.

Verified and unconfirmed

  • Verified: Secondary coverage reports that Harvard researchers found AI coding agents shifted the developer bottleneck to code review; the item itself supplies no primary paper, figures or quotations.
  • Not yet confirmed: The study’s methodology, sample size, models tested, publication venue and peer-review status; whether the shift is measured or inferred; and whether the effect holds outside the settings the researchers examined. Commenters also dispute how current the finding is, with one arguing such research lags practice by months.

Background

The finding traces to a Harvard working paper by economists Fiona Chen and James Stratton, who analyzed engineering telemetry supplied by Jellyfish, an engineering-analytics platform, covering roughly 718 firms and 700,000 workers between January 2021 and March 2026. That design measures pull-request throughput and review time across real repositories rather than running a controlled experiment, which is why the result is framed as a working paper about where developer effort moves — not as a benchmark of any specific coding agent.

Why it matters

For teams adopting coding agents, the constraint on throughput becomes review capacity rather than code production: one reviewer in the discussion, who reviews full time for a team, says agent-assisted merge requests carry more code and often reach for overdesigned solutions where a simpler one would do, which raises the cost of each review. That argues for capping agent-generated change size and requiring tests before review, rather than assuming faster code generation alone speeds delivery.

Community discussion

A developer with 25 years of experience says agents now write code roughly as well as they do, and find issues in code better still, so agents reviewing each other’s work produces acceptable quality. Others are less positive: a full-time reviewer reports receiving larger, overengineered merge requests from agents, another says reviewing AI-written code all day is unenjoyable, and one commenter argues university research on AI is dated relative to what practitioners encountered months earlier. These are individual opinions, not evidence of consensus.

Further reading

Comments

Comments

Comments use a GitHub account and are moderated. Comment rules

More in AI

See all