Primary sourceeric zakariasson
View sourceEvent ·Multi-source
Cursor 团队分享优化 Agent harness 省成本的提示词
Cursor 团队发布一段用于优化 LLM agent harness 的提示词,目标是按计费类型加权降低每个任务的 token 成本,同时不损失任务质量。核心原则包括:不要让模型自己省 token,有团队因此导致模型拒绝执行复杂任务;用工具的功能描述替代 DO NOT 类指令,有团队借此砍掉约三分之二的 system prompt。一个生产级编码 agent 经过一轮 prompt 精简、工具卸载、缓存布局调整、稀疏行号和子 agent 调优后,整体 token 成本降低约 7%。提示词还给出五步流程:先绘制 harness 结构并测量基线,再排序优化点。
Current read
Cursor 团队把内部学到的省 token 方法打包成一个提示词,直接喂给你的 agent 就能系统性降成本,还有真实案例数据。
11 sources73° AI heatLast updated 2026/9/23 20:12:00
Evidence
11 sourcesRelated source 1shao__meng
View sourceRelated source 2Cursor
View sourceRelated source 3Blognone
View sourceRelated source 4Lenny Rachitsky
View sourceRelated source 5AI Will
View sourceRelated source 6Claude
View sourceRelated source 7CoreWeave
View sourceRelated source 8Cloudflare Blog
View sourceRelated source 9arXiv: OpenAI
View sourceRelated source 10Fish Audio
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.