Two items are not enough for a factor analysis because they generate only a single correlation, but you certainly can combine them by taking an average, as you suggest.

Taking an average may be a workable proposition in this case.

Dear Benjamin Chua,

When you are conducting factor analysis you need  at least 3 items for a variables. because AMOS does not give results if you have less then 3 item. You can use the mean if you you are doing path analysis.  

All the best...!!

Benjamin,

A model with two indicators of a factor might be identified in the context of a larger structural model.

But as others have said, there's very little information there. If and only if the indicators are on the same metric, averaging is reasonable. But if you're running a larger SEM anyway, I would include the two-item factor in the larger model. If you run into identification problems and the indicators are on the same metric, then adding an identification constraint (e.g., that the two loadings are equal or both unit) is also reasonable.

Pat