Seven places Jev fits in an agent harness for routing and checks
The author is moving calls that only route, gate and check to Jev, at about 100ms and $0.042 per 1M input tokens, making 10,000 checks cost 84 cents.
What this article says
作者在自己的 agent harness 中实测 Jev,主张把那些只负责路由、门控和校验的调用从通用大模型迁移到 Jev。理由是这类判断不需要生成能力,只需要快速拍板,而 Jev 的延迟约 100 毫秒、输入定价每百万 token 0.042 美元,成本与延迟都远低于通用模型。作者算了一笔账:1 万次检查只需 84 美分,并强调自己核算了两遍。文章列出 Jev 在 agent 工作流中最合适的 7 个位置,核心结论是:agent 一个任务往往要经过几十次串行判断,如果每步都用通用 LLM,延迟和成本会不断累积,因此把纯决策环节交给专用决策模型是必要的工程优化。
I've been testing Jev in my agent harness.
I'm moving the calls that only route, gate and check to it: ~100 ms, $0.042 per 1M input tokens.
I wrote up the 7 places where it fits.
10,000 checks for 84 cents. I did the math twice. https://x.com/i/article/2108555620240547840