There is actually no rule. I would recommend the following steps to be able to model correctly one direction (assuming you have one and the tolerance is chosen):

1) Do plenty of lags (100) for whom the sum would be the full distance.

2) At this point there should be a lot of noise in the data but at the very least you should be able to tell more or less the distance at which the experimental variogram goes over the sill. Let's call that the distance to which you consider data to be correlated.

3) Now you keep increasing the size of your lags (and decreasing number, 10-15) to be able to have stronger means and thus (typically) less noise.

4) Once you've got strong enough means you should start to get (again this is common but not always) something that resembles a function or model.

5) You adjust the model with any number of structures you need.

A final note about your question. Notice that the size of the lags should usually be at the minimum of the distances between points for the direction you choose. You won't be gaining more information for smaller sizes only more empty lags.

Hope it helps.

Just to add one more point................usually in most of the cases, half of the maximum inter-station distance is enough for active lag distance. In most of the variogram modelling packages such as GS+ use it by default, although you can use the full distance as well. But active lag distance equal to half of the maximum interstation distance is enough and it is mentioned in many materials such as GS+ manual. Alternatively, you can find it in "Practical Geostatistics" by Isobel Clark published in 1979.

Use this technique and do perform your cross validation, hopefully, you will see the variogram models are perform well with satisfying all cross-validation criteria such as MSE, RMSE, ASE, RMSS. You can find all details of these statistics in ArcGIS Geostatistical Analyst extension manual.