📄 中文摘要
在并行运行多个 AI 编码代理的过程中,作者体验到了高效的短暂冲刺,但随之而来的却是混乱的代码管理。不同的代理在同一代码库中独立工作,导致相同功能的不同实现相互独立,彼此之间毫不知情。一个代理实现了动态模型发现,另一个代理在不同的工作区中也实现了类似功能,但使用了不同的类名,而第三个代理则在其功能中内联了自己的实现。最终,作者发现自己在三个不同的分支上拥有三个版本的相同概念,造成了代码的混乱和维护的困难。
📄 English Summary
Agentic Drift: It's Hard to Be Multiple Developers at Once
Running multiple AI coding agents in parallel can lead to bursts of productivity, but it also creates significant challenges in code management. Each agent works independently on the same codebase, resulting in multiple implementations of the same feature without awareness of each other. One agent added dynamic model discovery, while another, in a different workspace, implemented a similar feature with a different class name. A third agent required model listing and created its own version, leading to three distinct implementations across different branches. This situation illustrates the complexities and potential confusion that arise when managing multiple AI agents simultaneously.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等