Replacing agent decision nodes with Jev
The video argues agents shouldn't use generative models for decisions: Jev returns schema-guaranteed typed decisions such as which tool to use, whether to continue or stop, and when to escalate to a human, at about 1/200th the cost of frontier models.
There's a new AI model that doesn't generate text. it just decides. and that one difference is about to change how every AI agent is built. every agent you've used runs on the same architecture: action needed → call GPT or Claude → parse the text response → execute. slow. expensive. and one hallucinated tool call breaks the whole chain. Jev replaces that middle step: not "here's my answer in text". just - use this tool. stop here. escalate to human. execute now. typed decisions. schema-guaranteed. 0% structured output errors. what that means for agents: every decision node that used to be a frontier LLM call becomes a Jev call. which tool to use, whether to continue or stop, when to hand off to a human. the frontier LLM stays in the pipeline. it handles the steps that actually need generated text. Jev handles everything else. at 1/200th the cost. the agents we're building today are expensive because we're using content generators to make decisions. that's the wrong tool for the job. and now there's a right one.
This post quotes another post: x.com/p0lybender/status/2101765571096506841