一位软件开发者分享了过去18个月在AI辅助编程领域的实践演进历程，其工作模式从最初的Tab补全发展到Agentic编码，并最终演进为使用Harness管理多个AI Agent 1。在该实践中，作者构建了一个包含planner、worker、critic和promoter四个隔离角色的Agentic工作流 1。通过结合Cursor、Claude和Pi三个TUI工具，开发者能够共享统一的Harness体验 1。
在具体技术实现上，作者采用了prewalk技术，即使用前沿模型完成规划和首个任务，随后将后续工作交由其他模型处理，其中主要使用deepseek-v4-flash-0731处理维护和简单任务 1。这一方案成功将前沿模型Fable的使用量大幅降低了75%，从而帮助单人开发者实现了媲美大型团队的产出水准 1。此外，作者开发的一款将照片转换为矢量图形的相机应用产品，已通过poster-driver与Harness完成集成 1。该Harness的目录结构包含AGENTS.md、skills、extensions、artifacts和work等一级目录 1。作者还提及，在政府禁止使用Fable之后，大约有十万名开发者开始转向尝试中国模型 1。
A software developer has detailed their 18-month progression in AI-assisted programming, evolving from basic tab completion to agentic coding and managing multiple AI agents through a custom harness system 1. This agentic workflow is structured around four isolated roles: planner, worker, critic, and promoter 1. By integrating terminal user interface tools like Cursor, Claude, and Pi, the developer established a unified harness experience across all three platforms 1. The system utilizes a prewalk technique, where a frontier model handles planning and the initial task before delegating subsequent work to other models 1. For routine maintenance and simpler tasks, the developer primarily relies on the deepseek-v4-flash-0731 model 1. This methodology has decreased the usage of Fable, a frontier model, by 75%, enabling the solo developer to achieve the output capacity typically expected from a large team 1.
The developer's primary product is a camera application designed to convert photos into vector graphics, which integrates with the harness system via a poster-driver 1. The harness directory structure includes top-level directories such as AGENTS.md, skills, extensions, artifacts, and work 1. The author also noted that after the government banned Fable, approximately 100,000 developers shifted to experimenting with Chinese models 1.
评论
还没有评论，欢迎留下第一条。