{
  "id": 18760279,
  "title": "How we review 4x more AI code without reading every line",
  "url": "https://urgent.news/2026/11/07/how-we-review-4x-more-ai-code-without-reading-every-line",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-07T17:30:00.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/scale-ai-code-reviews/"
  },
  "original_language": "en",
  "account": "Code review doesn't have to mean reading every line of AI-generated code to trust it. Instead, focus on five key aspects: the intent behind changes, acceptance criteria, invariants that cannot be broken, decisions made during the process, and evidence confirming the code's functionality. At Aviator, implementing this approach enabled nine engineers to merge four times more pull requests, reduce the median time to merge from 2.1 days to 3.4 hours, and lower reverts by 40%. However, a human still signs off on each pull request. The key change was what that human reads. Review comments often consist of either established rules or personal decisions. While machines should handle rule checks consistently, decisions require human input as they are best addressed through discussion rather than within a diff. Code review efficiency faltered when transitioning from human-written AI-assisted code to AI-generated code. Weekly merged pull requests jumped from around 100 to over 400, with many reaching 640 lines, exceeding the commonly considered threshold for effective bug detection. Bug detection was never the primary goal of code review. Review was primarily about sharing knowledge and maintaining the team's understanding of the system, which is worth preserving. The issue arose when agents primarily wrote the code, making it the most vulnerable part. The most frequent criticism of code reviews is that they are outdated. We didn't discontinue reviewing altogether; instead, we stopped reviewing the code diff. The review agent analyzed the author's agent during the PR creation phase, identifying disagreements. In one recent PR, agents raised four disagreements and resolved three by stating the reasons. Only one remained for human evaluation, a naming dispute. Review agents also captured intent and acceptance criteria from the coding session, creating criteria based on engineer corrections and answers. While intent cannot be checked automatically, criteria can. The review agent mined a year's worth of review history, filtering out noise, classifying comments, and creating rules and invariants. For instance, \"Every config change writes an audit-log entry, only when something changed\" became an approved invariant covering specific configurations. This invariant is verified through a preview run and contributes to a model that checks invariants on every PR. The process takes 0.3 seconds and involves 19 deterministic invariants. Criteria are proven using various methods, including screenshots, request/response recordings, and file-and-line citations, depending on the criterion's nature. Finally, decisions should be debated.",
  "summary": "Do you need to read every line of AI-generated code to trust it? Not necessarily. Instead, check five things: the The post How we review 4x more AI code without reading every line appeared first on The New Stack .",
  "key_points": [
    "Focus on intent, acceptance criteria, invariants, decisions, and functionality evidence",
    "Aviator's approach enabled nine engineers to merge four times more pull requests",
    "Review agents captured intent, criteria, and created invariants for code verification"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}