{
  "id": 19850022,
  "title": "The Diff I Will Stop Reading",
  "url": "https://urgent.news/2026/11/13/the-diff-i-will-stop-reading",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-13T19:04:37.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/remuslazar/the-diff-i-will-stop-reading-2ig3"
  },
  "original_language": "en",
  "account": "In a recent private project working on an embedded device, the author experimented with an agent writing code. Initially, the author focused on reading small pull requests line by line, then shifted to reading only the tests, and eventually started asking questions about the design and potential issues. After checking the answers, they found them to be accurate every time. The author compares this experience to treating the agent similarly to a compiler and explores the reasons why the diff (changes) seems to be losing its value.\n\nFirstly, there is the matter of speed. The author notes that the number of changes in their repositories has more than doubled during the summer, while the median change remains around 35 added lines. Small steps help keep each change readable, but they do not necessarily make the total change more readable. When an agent produces a clean change in a few minutes, it takes the author ten minutes to read it properly, resulting in a clear arithmetic difference.\n\nSecondly, fluency plays a role. In a previous article, the author mentioned that their hands have gone numb, making it difficult to write code from an empty file anymore. The feeling of something being off in a diff before being able to explain it comes from having written thousands of them. Diff review relies on this feeling, which is fading as the author's fluency decreases. The author compares this to a two-thousand-line change that cannot be reviewed but only approved, which would also be true for a hundred and fifty lines.\n\nLastly, the diff was never where the problem broke. In a specific case, twenty clean changes to a deduplication step carried one wrong assumption for fourteen months. The author reread the change that should have caught the issue and found no objections. The diff review was the level where the author was best at, but the problem was not there. The compiler, as a trustworthy layer, is not the same as an agent, which fills gaps and makes decisions. The compiler is deterministic, while an agent does not reveal its decisions, making it unreliable.",
  "summary": "A compiler translates. An agent also fills the gaps. This autumn I worked on a private project, a small embedded device. Like the GStreamer patch , it was C. This time the agent wrote the code, with patterns I had never used. I started the way I would on a production system: small pull requests, each one read line by line. Then I noticed how good the code was, and I started reading only the…",
  "key_points": [
    "Increased speed of changes and decreased fluency in reading diffs contribute to the loss of value.",
    "Diff review relies on intuition, which is fading as the author's fluency decreases."
  ],
  "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."
}