六个月完全使用 AI Agent 编写代码的经验分享

exe.dev 的一位工程师近日分享了自己过去六个月完全使用 AI Agent 编写代码的实践经验 1。该工程师在 2024 年 2 月制定了一项规则，即不再手写代码，完全由 AI Agent 完成编码工作 1。在此期间，其工作模式从最初依赖 Copilot 和 Cursor 的自动补全，逐步过渡到利用 Claude Code 等 Agent 并行执行任务，最终转向构建工具以在隔离的虚拟机中管理多个 Agent 1。这一过程探讨了 Agent 驱动开发的工作流、安全隔离、代码审查与架构设计的重要性，体现了从“写代码”向“工程系统设计”的范式转变 1。

在模型能力方面，GPT-5.3 和 Opus 4.6 模型能够处理更大规模的变更且结果质量足够好 1。该工程师加入的 exe.dev 公司提供可在几秒内启动的 Linux 虚拟机 1。基于此，其构建了名为 botd 的工具来统一管理多个 Agent，支持移动端操作并保留所有对话记录，高峰期同时运行约 20 个虚拟机 1。此外，作者还构建了名为 Shelley 的编码 Agent 和名为 Athena 的部署监控 Agent 1。然而，botd 在本月发生崩溃，原因是它完全是 vibe coded 的，作者没有阅读其任何代码或关注其架构 1。在安全方面，红队 Agent 成功发现了作者以为已限制但仍可访问的开放网络路径 1。

The author, who joined exe.dev—a company providing Linux VMs that can boot in seconds—shared their experience of writing code exclusively with AI agents over the past six months 1. In February 2024, the author established a rule to stop writing code manually and rely entirely on AI agents 1. The workflow transitioned from initially depending on auto-completion tools like Copilot and Cursor to utilizing agents such as Claude Code for parallel task execution 1. Furthermore, the author noted that models like GPT-5.3 and Opus 4.6 are capable of handling larger-scale changes with sufficiently good results 1.

To manage multiple agents uniformly, the author built a tool named botd, which supports mobile operations and retains all conversation logs 1. At peak times, approximately 20 VMs were running simultaneously 1. The author also built a coding agent named Shelley and a deployment monitoring agent named Athena 1. Additionally, a red team agent successfully identified open network paths that the author had assumed were restricted but were still accessible 1.

However, the botd tool crashed this month because it was entirely vibe coded, meaning the author did not read any of its code or pay attention to its architecture 1. Through this six-month experiment, the author explored agent-driven development workflows, highlighting the importance of security isolation, code review, and architectural design 1. The experience underscored a paradigm shift from writing code to engineering system design 1.