Event ·Multi-source

Cursor 团队公开 Agent Harness 降本方法论

Cursor 团队成员 @ericzakariasson 分享了一套把 agent 降本当系统工程做的方法论,优化对象是 harness 本身而非模型行为。核心指标是按计费类型加权的“每完成任务”token 成本,一轮完整优化降低总成本约 7%。文中给出多个具体数字:移出静态上下文使工具描述 token 降 60%,MCP 集成工具相关会话总 token 降 46.9%,冷缓存未命中降 20%。多智能体场景用前沿规划器加便宜 worker,成本约为前沿模型全包的八分之一;简单轮次路由到便宜模型成本降 41–68%。

Current read

Cursor 团队把 agent 降本拆成了可上手的清单:砍提示词、工具卸载、缓存布局都有具体数字,做 AI 编程工具的能直接抄。

11 sources67° AI heatLast updated 2026/9/24 02:12:13

Evidence

11 sources
Related source 1eric zakariasson
View source
Related source 4Cloudflare Blog
View source
Related source 6Lenny Rachitsky
View source
Related source 7Fish Audio
View source
Related source 10Guillermo Rauch
View source

Conflict 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.