Grok Bot 工程师讲 agent 信任层，并对比 Jev 做路由决策
一场讲 800 PR/月背后信任层的 workshop 笔记：约束硬编码、路由仍靠前沿模型输出段落，而 Jev 返回带概率的类型化选项、每百万输入 $0.042。
SpaceXAI engineer behind Grok Bot just dropped a free workshop This 1-hour session on the trust layer behind an 800-PR month: 05:36 - her agents automerged 20 PRs while she slept 10:40 - no verification skill and you become the verifier, the bottleneck 24:01 - a judge agent of a different model cross-checks the score 35:04 - grok bot gives every agent its own identity 48:07 - ci checks turn the build red, rules and bugbot stay soft 57:13 - grok bot is the iMessage moment for everyone outside tech 800 PRs in 12 days. 600 of them refactoring the codebase so the agent stops needing her But every one of those constraints is hard-coded Directory the agent can't import from, a check that fails the build Every routing decision inside that team still goes back to a frontier model, which answers with a paragraph Chatbot used as a switch statement Jev returns typed options plus a probability, $0.042 per million input tokens, and never writes a sentence LLM creates the work -> Jev decides what happens next -> code enforces it Save this before you open another agent tab
这条是引用着另一条写的：「Jev 把路由打分验证做成决策层」