WHAT NEW IN X10.29 THU

X 上的 AI 合集档案 · 22 个合集 · 60,636 条

jevVIDEO

Jev 1.13 当 agent 决策芯片:万次决策 $0.42

作者演示用 Jev 1.13 把 agent 的分支判断从 LLM 里拆出来:读状态、选 Choice/Score/Yes-No、低置信度转人工,并给出 token 成本账和一个把 Claude 会话工具调用从近 1M token 压到 86K 的插件。

Dark Factor
@DarkFactorr6 天前收录原推发布于 2026.10.23
Dark Factor @DarkFactorr原推发布于 2026.10.23
在 X 上打开原推

this is f*cking insane, Jev 1.13 works like a decision chip for ur agents. state goes in, a typed answer with a probability comes out, and 10,000 decisions cost $0.42 most agents send every fork to the same LLM that does the work, so which worker goes next or is this ready becomes another model call before anything useful happens Jev by TypeSafe takes those forks off the LLM: > read: the goal, the work done and the workers available right now > choose: Choice for a route, Score for a rating, Noul for yes or no > check: anything under ur confidence threshold goes to review > hand off: code executes the decision and queues the next job • what the token math actually looks like > $0.042 per million input tokens, nothing for output > 1,018 research papers classified for $0.08, 256ms each > 500 emails sorted in seconds for 3.5 cents • where the autonomy sits, and where it stops > options rebuilt from what exists now, not yesterday's menu > confidence is not accuracy, uncertain calls stop at review > publishing stays behind a separate approval and the part that broke my brain: one plugin used Jev to score every tool call in a Claude session nearly 1M tokens dropped to 86K in about a second the expense was never one decision it was every decision that sent a worker down the wrong branch ↓

这条是引用着另一条写的:「长文:如何搭好 Jev 驱动的 agent 技术栈」

x.com/darkfactorr/status/2113348618375561370采集于 2026.10.23 10:06