There are complicated ways to specify such an effect in SEM, but the reality is that there is only one correlation between the two variables, so any attempt to turn that into two separate paths is a dubious exercise.
Lait, you would need single predictors only affecting one of the dependent variables which are supposed to have a reciprocal effect. If all independent variables would affect all dependent variables simultaneously, then the unique parts of the dependent variables that form the reciprocal effect could not be separated.
Do not test each path in SEM separately. Of course, it depends on what you are studying (for example if you researching about human intentions then you cannot test each cognitive element of intentions separately because we are talking about a complex human mind that needs SEM analysis to test everything at the same time instead of using regression).
The real and true significance is when you achieve the p-value of less than 0.05 by testing your entire model in SEM software at the same time. If you are using the software AMOS then use the bootstrap method in order to generate p-values for each path or line at the same time.
Search for an a journal article called "Beyond Baron and Kenny: Statistical
Mediation Analysis in the New Millennium" by Andrew Hayes (2009). It explains the bootstrap method.