Jev 跑 Grok 多智能体交易决策台：93.6 倍提速、444.6 倍降本
作者用八个 Grok agent 审查一笔 5375 美元候选订单，Jev 以受限格式路由最终状态，并质疑 TypeSafe 官方性能数字需独立验证。
93.6x faster. 444.6x cheaper. Those are TypeSafe's own headline numbers for Jev, and they are exactly the kind of numbers that should trigger skepticism before excitement. The company says the result comes from System One workflow evaluations. It also says these figures are probably at the high end of real-world gains. Published pricing is $42 per billion input tokens, with output listed as too cheap to meter. Now put that claim inside a system that makes thousands of small decisions. In this concept desk, eight Grok agents investigate a $5,375 candidate order. Risk and liquidity reduce it to $1,255. JEV routes the final state in a constrained format. When the 9.9-second freshness window expires, FENCE forces the paper order to $0. Even a 10× real-world improvement would change how many checks an autonomous workflow can afford before acting. At 193.6×, the architecture starts looking less like a chatbot and more like a new software primitive. The benchmark still needs independent testing. The economic question is already interesting.
这条是引用着另一条写的：x.com/xmyttle/status/2102028507433836576