Jev 在编码 agent 里做继续/停止/复核决策
一个 agent 跑出 240 个测试用例、19 个失败，作者用 Jev 返回 Choice/Score/概率来决定 harness 何时继续、停止或请人复核。
THE MOST IMPORTANT MODEL IN THIS AGENT DIDN’T WRITE A SINGLE LINE OF CODE. Jev’s job was to help decide what happens next. the build in this screenshot claims 55 minutes and $4.92. but the number that caught my eye was buried lower: 240 test cases. 221 passed. 19 failed. most agent demos stop at “it built the thing.” this one shows why the harness matters. a coding model builds. tests expose what broke. Jev returns bounded decisions a choice, a score or a yes/no probability. the harness decides when to continue, stop or ask for review. Jev doesn’t magically make the 19 failures disappear. it gives the system a way to make decisions about them without asking another chatbot to write “looks good.” the cheap build is the hook. the visible failures are the engineering. I broke down where Jev fits inside the agent stack below ↓
这条是引用着另一条写的：x.com/mr_kozh/status/2103131220884529370