Jev 用于 computer use 与浏览器代理的降本尝试
作者把工具选择、目标元素、操作类型等封闭集判断交给 Jev，减少 LLM 上下文与成本，仍在实测未下结论。
I found Jev (@TypeSafe) very promising for computer use and browser agents, especially for lower latency and LLM cost on tasks that are purely about choosing from what's on screen. Computer and browser use kept getting better with every new model, but it all rested on the model being smart enough to take a screenshot, reason about the image, and move forward, and every step burned more tokens. At this point, I eat through Astra credits in hours. Using LLMs for heavy and scalable CUA was too expensive. Tool selection, target element, operation type, and state are all classifications over a closed set and should have been treated as one when handling them via agents, but so far I "let the LLM handle it" because it was easy. But with Jev, it seems equally easy for me to switch to a hybrid setup and reduce the burden on my LLM's context and cost. I'm still playing around and don't want to add to the "hype," but I've genuinely felt that Jev has helped my agents find a new way to delegate.