Great Models Think Alike: Improving Model Reliability via Inter-Model Latent Agreement