single arrow stands for causal relation and double arrow stands for correlation.
normally, you would run a preliminary model with all the latent variables correlated.
if the model is identified, that means your causal model will be identifed, too.
as to the causal relationship, that is, the direction of the arrows need to be guided by theories and previous studies.
you can also use modification index as help to add or remove the paths, but still the model needs to be according to theories instead of SEM program.