AgentKit 接入 Jev 做可插拔语义决策层
作者实测两周，把技能路由、复杂度打分、评审分流等小决策交给 TypeSafe Jev 等决策模型，且默认关闭、需用户显式开启。
Been testing a small experiment inside AgentKit for the last 2 weeks: What if the agent DOESN’T ask a giant LLM every time it needs to make a tiny decision? Turns out… yeah, that makes a lot of sense 😁 So AgentKit now has a pluggable semantic decision layer. Today: TypeSafe Jev Cloudflare Clef Cloudflare Clef-flash I’m using it for stuff like skill routing, complexity scoring, review triage, test impact, workflow decisions, etc. AgentKit is my toolkit for building production-ish AI agent workflows, so I’m intentionally paranoid about this stuff: Decision models can advise. They do NOT magically gain authority. Everything is opt-in at user scope, credentials stay user-level, and a cloned repo cannot turn it on for you. Still dogfooding both providers. Let them fight 💪