Suppose there are two variables both of which are endogeneized in a structural model. For instance, people's retirement decision ($X_1$) and saving behavior ($X_2$) are endogeneous because they are the results of the utility optimization procedure given all other exogeneous variables. In other words, given all exogeneous values, we can find out the level of $X_1$ and $X_2$. If we believe in this logic, I would say there is no causal relation between $X_1$ and $X_2$ as the relation is determined purely by other confounders. Am I correct? If yes, causal relations from many reduced form estimation are then hard to be convicing because many variables are endogeneous in some sort of "structural model".
I am trying to identify the causal relation between the number of children and health utilization but I meanwhile realize that both two variables are the by-product of a given utility optimization problem.
That is why economist moved on from the conditioning on all the confounding variables to design-based methods such as instrumental variables, regression discontinuity, etc. As long as you can find something that moves your variable of interest independently, you can claim (with a lot of additional work) that the estimated effects are causal.
In your example you can think of some regional policies aimed at fertility (that would create exogenous variation in number of children) that do not affect health. Then, instrumenting with these policies, would potentially allow you to avoid the endogeneity problem that you are describing.