用 Jev 三类原语写编码 agent 的决策循环论文
作者把 Jev 思路整理成论文：YES/NO 做门禁、CHOICE 做路由、SCORE 评风险质量，让前沿模型只处理真正需要推理的部分。
i still don't understand why everyone is letting coding agents burn frontier-model tokens on every tiny f**king decision JEV fixes exactly this i turned the JEV engineering approach into a research paper on how to build coding agents around typed decision loops instead of asking the big model to reason about everything, JEV handles the small repeated decisions with 3 primitives: YES / NO for bounded gates CHOICE for routing SCORE for risk, quality and relevance the expensive model stays focused on the work that actually needs reasoning the paper breaks down the full loop: routing, context, browser state, code review, verification, human approval and when the frontier model should actually be called one useful coding agent isn't just one giant model it's a system that knows when the expensive brain actually needs to think.
这条是引用着另一条写的：「把 Jev 放在 Claude Code 前面做成本路由」