Handing the agent loop's judgment calls to Jev as typed decisions
After logging an agent run where most model calls produced no text, the author split the loop's decision points out to Jev for typed answers with probabilities, letting confidence decide between automating, gathering more evidence or escalating to a human.
i logged every single model call inside one agent run and most of them never produced a single word the bill was never the real story, the real story was how many branches in that loop were asking a frontier llm to do a job it was never built for jev engineering: stop letting an llm negotiate its own permissions here's the pass it went through: ~ map, find every decision point in the loop before touching a line of code ~ class, sort each one into generation, semantic judgment or a plain hard rule ~ swap, hand the judgment calls to jev and get typed answers with probabilities instead of prose ~ route, let confidence decide between automating, gathering more evidence or escalating to a human ~ lock, keep permissions and budgets in code where nothing probabilistic gets to negotiate them people picture an agent as one brain doing everything, ask, answer, act, repeat this one pulls the brain apart instead, the llm creates, jev decides, code enforces, and a confidence score never gets to become permission on its own the routing is the part that actually matters, a 0.51 and a 0.99 stop getting treated the same the second probability is allowed to change the path full split, the thresholds and what each layer costs is in the breakdown below
This post quotes another post: x.com/cryptowluha/status/2103600863868604579