Putting Jev between every model call stops the agent from guessing
The author uses GPT-6 Astra for reasoning and Jev for routing decisions at each state fork, ran a week of shadow logs with 91% agreement with human judgment, and found full-LLM decision routing was 193x slower and 444x more expensive.
still can't believe this is real. i put Jev between every model call and my agent stopped guessing state, decision, action, verification, next state - most agent stacks are still missing one link in that chain, and it's not the model a team benchmarking their own stack found the gap: routing every decision back through a full LLM call ran 193x slower and 444x more expensive than routing it through a dedicated decision layer instead. the model was never the bottleneck. the fork between one state and the next one was i put GPT-6 Astra behind the same fork for a week: reason with Astra, decide with Jev, execute with tools, update state, decide again. same loop, on repeat this is Jev Engineering, the layer that turns every decision fork into something you can benchmark and verify on its own, $0 on top of the plan you already pay for: - Astra reasons once per state, not once per decision - it stops re-explaining itself every loop - Jev owns the fork: given the same state twice, it routes the same way twice, which a model alone won't promise - tools execute only the route Jev picked, so a bad call fails at one step instead of cascading through the whole session - every fork gets logged separately from the reasoning that led to it, so you can replay just the decisions - stay shadow mode first - log every route without acting on it - flip it live only once the logs agree with what you'd have picked by hand a week of shadow logs matched my own calls on 91% of forks. the other 9% is where i still step in the catch: shadow mode alone doesn't save anything, you're just watching. the savings only start once you flip it live, and that's the point most people never get past every agent still routing every fork through a full model call is paying for a decision it already made yesterday bookmark this, the full state-decision-action loop and the exact config that got mine to 91% is in the piece below ↓
This post quotes another post: x.com/bl888m_eth/status/2100609025243324849