Modifications indices should be used as a guide and not the sole basis for adding paths to a model. The researcher should strongly consider whether adding certain paths is theoretically defensible in addition to empirically defensible. As I think you are using MLR as your estimator given previous posts, you should examine the mi.scaled column. Values over the critical value for chi square on 1 degree of freedom indicate that adding the path in the output MAY significantly improve model fit. This does not guarantee that doing so will and you will have to check each new model against the previous one that does not include the path you just added using the anova() function. Using modification indices is therefore an iterative process. Start with the largest, theoretically defensible mi add it to the model, check to see if the model fit significantly inproved, then request a new round of modification indices. Continue this process until you a) run out of MI's that suggest significant improvement or b) you run out of paths to add that are theoretically defensible.