ClickHouse 的 Agentic 编码实践
ClickHouse 团队分享了他们在 C++ 代码库中使用 Agentic 编码的实践经验。
文章指出，尽管对 AI 编码代理的看法存在争议，但 ClickHouse 通过使用 Anthropic 的 Claude Code 等工具，已经将代理应用于性能测试、代码规范和 bug 修复等任务，显著提高了开发效率。
文章强调了 CLI 代理的优势，并指出随着 Claude Opus 4.5 等模型的出现，AI 代理在大型代码库中的应用潜力巨大。
未来，Agentic 编码将成为提升软件开发效率的关键，但需要避免强制推广，并持续关注工具的成熟度和安全性。
查看原文开头（英文 · 仅前 3 段）
First coding models and agents are just one year old, and today there are polarized opinions about the use of agentic coding in practice. Some people will tell you that agents will take over all our jobs, and some will tell you that coding agents are totally useless. Some people hate AI for reasons, and there are people who have long gone into AI-psychosis. And if you read the news, it does not help either: every day it is a kaleidoscope of new frontier models, more advanced tools, new research and breakthroughs, tremendous results on benchmarks, and at the same time, low-quality code, security vulnerabilities, studies showing negative economic impact, and modest results of autonomous agents on real jobs. In many companies, leadership tries to mandate the use of AI, while employees feel confused and insecure.
I want to avoid this confusion and make it clearer. We use coding agents in ClickHouse, and they are a great tool for certain scenarios.
Note: I don't use AI for writing texts, because I don't like it. I write texts very slowly, but it's my approach.
※ 出于版权考虑，仅引用前 3 段。完整内容请阅读原文。