用 Jev 六问复盘,把结账 agent 成功率从 12% 提到 88%
同一模型不改权重,50 次对抗式结账测试中让 Jev 在每次失败后问 6 个结构化问题,保留 8 条规则、丢弃 16 次噪声失败,29 秒 0.02 美元。
I just built a Jev X Self-Improving Agent. 12% → 88% success. 29 seconds. $0.02. Zero fine-tuning. Same agent. Same model weights. 50 attempts at a hostile checkout. after every failure Jev asks 6 typed questions: which step broke, what was the root cause, will it happen again or was it noise. 8 rules kept. 16 failures thrown away as noise. ignore the green "BUY NOW" banner - +16 pts captcha → audio challenge - +12 pts the agent was never bad at checkout. it was bad at not clicking ads. Read this article on Jev Engineering below and turn your ideas into reality.
这条是引用着另一条写的:x.com/0xricker/status/2102396090012115030