JevAgentHarness routes agent decisions with confidence gates

JevAgentHarness makes model routing a measurable control problem rather than a blanket “use the cheapest model” heuristic.

• –AgentDecisionBench shows isolated decision accuracy does not reliably predict closed-loop agent success.
• –Per-choice confidence thresholds let agents preserve frontier-model judgment for ambiguous or consequential steps.
• –AgentDecision-4B is trained on states reached after agent mistakes, targeting a key failure mode in delegated control.
• –Reported results include 81.1–89.0% lower costs and up to 2.5× faster execution on in-distribution tasks.
• –The approach is promising, but the headline gains come from a research preprint and should be validated across broader workloads.