Architect's five counterintuitive agent-coding rules: throw away code you can't read
InfoQ 中文 AI&大模型 translated a podcast interview in which chief architect Carlos Kelkboom told host Patrick Akil five counterintuitive rules for agentic coding: build the code skeleton yourself, bar the main agent from editing files directly, discard generated code you cannot recognize, let agents finish larger tasks in one pass, and run several agents in parallel then merge only the good parts. He described familiar failures, including agents quietly replacing deployment scripts, the same feature implemented five times, and patching rather than restructuring when architecture needs to change. The two also discussed shrinking PR debate, how AI-generated code should be reviewed, and what happens to engineering judgment and junior-developer training as code gets cheaper.
The report frames these views as a critique of current agent-assisted development practice, arguing code can be generated faster but decisions about why a project exists, who it serves and how it evolves still need human judgment.

翻译|林绮蓓
编辑|蔡芳芳
“我已经很久没好好看过代码了。”在 AI 编程工具越来越强的今天,这句话开始成为一些开发者描述自己工作方式的方式。需求交给 Agent,代码由它生成,测试和审查还可以交给其他 Agent。软件似乎从来没有这么容易做出来过。
但另一种情况也越来越常见:项目启动得很快,做了几轮后却很难继续。Agent 悄悄换掉原有部署脚本,同一个功能在代码库里被实现了五遍;一旦需要大幅调整架构,它往往选择继续打补丁。开发者拥有了大量能运行的代码,却未必还看得清系统究竟是怎么搭起来的。
在首席架构师 Carlos Kelkboom 看来,问题已经延伸到人与人的协作。开发者遇到问题越来越习惯先问 Agent,PR 下的讨论逐渐减少,与用户沟通、理解产品需求、反复思考架构的过程也更容易被跳过。代码可以更快生成,但一个项目为什么做、为谁做、以后怎样演进,仍需要有人作出判断。