把 agent 路由层换成 Jev 解决生产失败
同一 agent 任务在脚本化 demo 里 41 秒通过，在真实仓库里却陷入 8 个子 agent、12 个工具、10 次重试的循环；作者把路由层交给 Jev 逐步选工具、第 3 次重试即止损、只在状态真正模糊时找人，真实仓库下能跑完。
a chinese dev just used JEV to kill the "demo works, prod fails" problem every agent builder is quietly dying from run A: scripted agent, task is "add pagination to /orders", 3 steps, 14/14 tests pass, done in 41 seconds. run B: same agent, same task, no fixtures, real repo. 8 sub-agents, 12 tools, 10 retries, 780 frames later. TESTS looping, DOCS stalled, REVIEW stalled, APPROVAL waiting on a human. internal adoption 97.9%. external adoption 0.7%. that gap is where every agent demo on this planet dies. the first version let the main agent pick its own route through 12 tools and ask for human approval every third step. then he moved the routing layer to Jev. now Jev picks the next tool at every step, kills loops at retry 3, and only escalates to a human when the state is actually ambiguous. no fixtures. no scripted path. real repo. the agent finishes. I pulled the exact routing policy he used and it reads less like a prompt and more like an on-call runbook. full breakdown in the article below ↓
这条是引用着另一条写的：x.com/0xricker/status/2102396090012115030