Wouldn't it make sense to decouple the orchestration later from the authoring layer for the dags? That way you could solve the authoring problems separately from the lower level orchestration problems. We're trying this over at ZenML (https://zenml.io) but have yet to get feedback