Reviews got faster and thinner after agents arrived, and in repos without agents too
A difference-in-differences study of 669 developers finds agent adoption tied to longer PR descriptions and more testing. The slide in review thoroughness shows up in control repositories as well.
Source: Does Working with AI Agents Change How Developers Code, Test, and Review?
Muhammad Hamza, Dominik Siemon, and Wardah Naeem Awan (ESEM 2026) test the automation-complacency worry directly: once agents start opening pull requests, do developers coast? They follow 669 developers across 103 repositories in the AIDev dataset before and after agent adoption, on 11 metrics covering coding effort, testing, and review. To separate agent effects from whatever else changed, they add a difference-in-differences comparison against 228 control developers in repositories with no agent activity.
The one large agent-linked effect is documentation: PR descriptions got 76.8% longer (r = 0.561). Testing improved modestly on all three measures. Coding effort held steady. Review did change, becoming faster, less detailed, and more permissive, but the same drift appears in the control repositories, so the authors attribute it to broader shifts in how software gets made rather than to agents.
Why it matters for code review
This complicates the easy story that agents make reviewers lazy. On this data, agent adopters wrote more context and more tests, not less. The thinning of review is real, but it is happening across the ecosystem, including on teams that never turned an agent on.
That is arguably worse news for anyone scaling continuous agent output. If review depth is eroding everywhere, the volume from always-on agents lands on a review culture that was already getting quicker and shallower. The paper's design is also a useful reminder for every before-and-after chart in this space: without a control group, a decline in review rigor gets blamed on the newest tool by default.