论文扫描 Jev 早期生态：各领域用法与关注度错配

Francesco @francescoinweb3原推发布于 2026.09.27

在 X 上打开原推

if you're picking where to build with Jev, this paper is basically a treasure map. someone scraped the entire early Jev ecosystem and plotted what people actually do with it vs where the attention is. two things jump out > first - Jev isn't one tool, its job changes by domain. in interface agents and simulation, "action selection" is the main use (47% and 52% of decisions). in routing, a quarter of all decisions are the model picking which model or tool to call next (27%, vs just 6% in software engineering). > but across six of eight categories, the top job is plain attribute judgment - "is this thing what it claims to be." same model, completely different role depending on where you drop it. > second - and this is the opportunity - attention is wildly lopsided. Routing & Automation and Interface Agents are only 19.6% of all Jev projects but pull 63% of the GitHub stars. > the biggest categories by volume, Content & Expert Tasks and Search & Memory, have ~387 and ~381 projects but average just 31 and 37 stars each. > the sharpest line in the whole thing: Routing and Simulation have almost identical project counts (250 vs 252) and average 364 vs 8 stars. same amount of work, 45x the attention. takeaway for a builder: the crowd is piling into routing. the big, under-starred categories - search, memory, expert tasks - are where the room actually is

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」