Beyond Fluency: Toward Reliable Trajectories in Agentic IR