Moving yes/no decisions in agent loops from LLMs to Jev
The author ran Jev across every live agent loop and found many yes/no judgments were being outsourced to generative models, so he split operations into generation, decision, and hard rules, letting Jev return typed answers with probabilities for the decisions.
i ran jev across every agent loop i had live and found out most of my frontier model calls were answering yes or no questions worth exactly $0 turns out the token spend was never the real problem, the real problem was how many tiny decisions inside each agent got quietly outsourced to a generative model that never needed to write a sentence jev engineering: stop paying an llm to answer yes or no here's the breakdown it forced: ~ split, pull every operation apart into three buckets, generation, decision, or a hard rule ~ count, log each yes/no, route and score as its own event instead of folding it into one average ~ type, hand the decisions to jev and get back typed answers with probabilities, not paragraphs ~ gate, run every tool call through a risk check before code lets it fire ~ trace, keep the full probability spread and threshold sitting next to whatever action it triggered people assume an agent is just one llm thinking out loud, one loop, everything tangled together, nothing sorted this setup sorts it instead, the llm keeps generation, jev owns judgment, code owns enforcement, and none of the three can pull rank on the others that separation is exactly what breaks the latency ceiling, route every branch through a frontier model and the whole thing stalls before the task even finishes confidence isn't uniform across these decisions either, so different parts of the same agent end up needing completely different routing at that point you're basically looking at real system architecture, even though the whole thing still counts as one agent under any definition people use today the boundary between a prompt and an architecture existed long before anyone tried to measure it, and it sits right on generation full cost and routing math for every layer is in the breakdown below
This post quotes another post: x.com/cryptowluha/status/2103600863868604579