{
  "id": 21078578,
  "title": "If we stop reading the code, what do we read?",
  "url": "https://urgent.news/2026/11/21/if-we-stop-reading-the-code-what-do-we-read",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-21T17:23:05.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/ai-agent-code-review/"
  },
  "original_language": "en",
  "account": "In an era where code agents can write and open pull requests at breakneck speeds, humans are finding it increasingly difficult to keep up with reading every line of code. Robert \"Uncle Bob\" Martin, a renowned software engineer and author, concedes that he rarely sees 95% of the code his agents generate. The speed of code agents is outpacing human reading capabilities, and most teams have long abandoned line-by-line reading.\n\nRather than completely abandoning code reading, Uncle Bob suggests focusing on high-level supervision. AI agents provide a visual representation of the code, allowing engineers to expand, collapse, and drill into specific parts. This tool enables them to identify issues, such as functions with high CRAP scores (Change Risk Anti-Patterns), low mutation-testing scores, and problematic dependencies. Engineers can also simulate what-if scenarios and review proposed structures before any changes are made to the source code.\n\nWhen dealing with a team, the reviewer should pay attention to several aspects when approving changes. They should examine the intent of the change, the acceptance criteria, and the architectural, behavioral, and technical decisions that were made. Intent is captured through prompts and back-and-forth with the agent, while acceptance criteria consist of specific change requirements and team-wide invariants. Decisions made during the process, such as library selections, data models, and fallback strategies, should also be reviewed.\n\nUltimately, the goal is to maintain a mental model of the codebase, even if engineers can't realistically read every line of AI-generated code. By focusing on intent, acceptance criteria, and decisions, engineers can ensure they understand the system's workings and the reasoning behind architectural choices, without being overwhelmed by the sheer volume of code.",
  "summary": "“Human beings are too slow when dealing with code. So if we’re going to move at the speed of the The post If we stop reading the code, what do we read? appeared first on The New Stack .",
  "key_points": [
    "Humans struggle to read all AI-generated code due to its speed",
    "Focus on high-level supervision with visual code representation",
    "Review intent, acceptance criteria, and architectural decisions"
  ],
  "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."
}