@Sumanth_077 on JevAgentHarness: Jev gates tool/arg/done calls; AgentDecision-4B cuts cost 81–89% in-loop.

Agent write-up with diagram. LLM agents make many small decisions between reasoning steps; JevAgentHarness routes them to a smaller decision model when confidence clears a trajectory-learned threshold, else falls back to the LLM. Standalone decision accuracy was not enough (best isolated model only 69.5% full-task success; naive delegation −19 pts). Training AgentDecision-4B on states after mistakes cut cost 81.1–89% and sped runs up to 2.5× without losing ≥3 pts success; on shifted tasks it beat Jev (91.5% vs 82%). Harness, benchmark, data, and models are open source. For agent builders putting decision models inside the loop.