围绕人工智能生成代码是否需要逐行阅读，编程界近日陷入激烈争论。[1]HashiCorp联合创始人Mitchell Hashimoto发推文主张"I read the code"，获得近83万次浏览，强调开发者应当审查代码；[1]而资深开发者Uncle Bob则提出对立观点，称"完全不去读Agent写出来的任何代码"，该言论获得超480万次浏览。[1]这场论战反映了AI时代代码生成成本下降后，开发工作方式的根本转变。[1]
Uncle Bob的核心主张是用约束和验证体系替代人工审查。[1]他提出的策略包括设置单元测试、Gherkin测试、QA流程、质量指标和变异测试等严格机制，而非逐行审查代码。[1]Uncle Bob认为代码质量仍然至关重要，混乱的代码会降低AI Agent本身的工作效率。[1]
对此，开发者Cindy Sridharan表达了担忧，指出"每当听见有人说代码全是Claude写的、自己不知道工作原理，我就认定这个人根本没有能力调试这些代码"。[1]另一位开发者Christine Lemmer-Webber提出"Vibe滑坡"现象：随着生成速度的增加，代码审查逐步被放松。[1]而Theo创始人则持不同看法，认为大多数工程师的代码阅读比例过高，应当从AI生成代码中获得更多价值。[1]
A significant debate has emerged within the development community regarding whether programmers should carefully review code produced by artificial intelligence systems. Mitchell Hashimoto, co-founder of HashiCorp, posted on social media advocating for the position "I read the code," emphasizing developer responsibility for understanding generated code, garnering approximately 830,000 views as of July 3 [1]. This stance prompted a contrasting response from veteran developer Robert C. Martin, commonly known as Uncle Bob, who declared "I don't read any code written by agents at all," a post that achieved over 4.8 million views [1].
Uncle Bob's alternative approach centers on establishing rigorous safeguards rather than relying on manual code review [1]. His proposed framework includes implementing comprehensive unit tests, Gherkin-based testing, quality assurance processes, quality metrics, and mutation testing to verify code integrity [1]. This methodology reflects a fundamental shift in development practices as artificial intelligence reduces the cost of code generation [1]. Other industry voices have contributed to the discussion, with Cindy Sridharan arguing that developers claiming "all code is written by Claude and I don't know how it works" demonstrate an inability to debug such code [1]. Christine Lemmer-Webber identified what she termed a "vibe slope" phenomenon, where increased development speed correlates with reduced code review rigor [1]. Despite these differing perspectives, Uncle Bob maintains that code quality remains important, as he believes poorly structured code diminishes the efficiency of AI agents themselves [1]. Theo's founder has suggested that most engineers spend too much time reading code and should extract greater value from AI-generated solutions [1].
