JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs

JevAgentHarness is designed to let decision models share control with LLMs in agentic tasks, reducing the need for large LLM calls at every step. The approach uses AgentDecisionBench and DecisionGym to evaluate decision models, fitting confidence thresholds on the agent's own trajectories to delegate tool choices.