well. we are using it from our code and the trains dashboard works just fine. But this is without the agent. Moreover, Kubeflow uses docker images for the training pipelines, there for trains cannot capture the diff