Jev + GrokBot 做 agent 路由省预算
作者用 Jev 在昂贵工具调用前做小路由决策（复用缓存、停止重试、限制调研、派子 agent），GrokBot 负责执行；本地测试受限调研 4.125 秒 vs 不限 53.803 秒，建议先跑 shadow 模式对比。
Jev + GrokBot might be the best AI agent setup Ive come across because it tackles the part that quietly burns your budget unnecessary searches browser calls you didnt need the same failed approach tried again Jev makes a small routing decision before the expensive work starts reuse a cached result stop the retry limit the research bring in a subagent GrokBot handles the execution In one local test shared by the author the capped research run took 4.125 seconds compared with 53.803 seconds uncapped It fetched fewer pages so that isnt a universal speed guarantee But the approach is worth testing Start in shadow mode compare the decisions with your existing workflow check what it skips and what it misses The biggest upgrade to your agent might be teaching it when the next tool call isnt worth making Repo in the reply