用 Jev 替换 agent 决策层省 100 倍成本的 6 步
审计 LLM 调用区分「写」与「判」，Gemini 写、Jev 判、代码执行，用 Choice/Score/Null 加 rubric 与批量门控。
Jev Engineering cuts AI agent costs by 100x on Google's stack your agent spends 90% of its budget deciding, not thinking. Jev makes those decisions for almost nothing here's how to rebuild your agent's brain in 6 steps: step 1 → audit every LLM call. log each one and tag it: "writes" (drafts, summaries, code) or "decides" (routing, scoring, yes/no). most agents run 70%+ decisions step 2 → split the architecture. Gemini writes. Jev decides. code acts. each job goes to the component built for it step 3 → structure your decisions. Jev takes three types - Choice (pick from options), Score (rate on a scale), Noul (yes or no as probability). no open-ended prompts step 4 → build rubrics, not prompts. every decision gets a JSON rubric with definitions, boundaries, and examples. version-controlled, testable, no hallucinated options step 5 → batch and gate. send all questions in one call (12x cheaper than sequential). set confidence thresholds - high confidence auto-acts, low confidence escalates to Gemini step 6 → measure cost per completed task, not cost per call. a cheaper decision that breaks the workflow costs more than the expensive one that worked the math: 10,000 daily decisions - $35,000/month on frontier models. $120/month on Jev. same accuracy, 193x faster bookmark this. your agent bill will thank you ↓
这条是引用着另一条写的：x.com/choopyplug1/status/2088973320964215253