Engineers have stopped reviewing PRs
SMRTR summary
Seventy-nine percent of AI-generated pull requests are reviewed by the same developer who wrote them. That's the finding from a July 2026 LeadDev analysis of more than 25,000 agent-generated pull requests, and it points to a quiet crisis in how software teams are managing AI-written code.
The problem isn't the code itself. It's what gets lost when a reviewer only sees the finished diff, and not the messy session that produced it: the original prompt, the mid-run corrections, the paths the agent abandoned, the warnings someone waved past under deadline pressure.
A new framework called "session review" argues that AI code review has to go deeper than the diff. It means preserving the original prompt, keeping the transcript retrievable, and making a second human reviewer the default, not a courtesy.
The core insight is almost unsettlingly simple: the decisions that determine code quality happen during the agent's run. By the time you're reading the diff, the most important part of the story is already gone.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article