用 Jev 做 agent 决策层：状态、提问、门控、行动
作者拆解 Jev 作为 agent 决策层的架构：状态输入、拆分提问、结构化选择与评分、代码门控、执行并回传结果，避免重复失败动作。
YOUR AI AGENT CAN WRITE A PERFECT PLAN AND STILL MAKE A STUPID NEXT MOVE. Wrong tool. Wrong queue. The same failed action for the fourth time. A better paragraph doesn't fix any of that. That's the part of Jev worth paying attention to: the decisions between the impressive answers. Which worker goes next? Is there enough information to act? Does this case need a person? The useful part of this architecture is the handoff. Each stage has one job: → STATE: give the system the current task, available tools and the result of its last action. If the tool failed, that failure belongs in the next decision. → QUESTIONS: separate the calls. Choose the next worker. Score urgency. Check whether information is missing. Several narrow questions can share one state in a single Jev request. → JEV: return structured choices, scores or yes/no probabilities. The coding model can keep writing; the router gets an answer software can use. → GATE: code checks the answer against the workflow's rules. Uncertain cases go to review. A confident model still doesn't grant itself permission. → ACTION: run the mapped tool, record what actually happened, and feed that result into the next turn. That last step matters. Without feedback, your agent can keep choosing an action that already failed. With a record, you can inspect exactly where the loop went wrong. The clip illustrates the architecture; it isn't a live-run benchmark. Jev can still make a bad call, so test the routing on cases you already understand. The Jev article below breaks down that decision layer. Look at your own agent's last run: how many model calls actually needed to write anything?
这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355