Why doesn't the transformer use positional encoding in every layer?

Question

Positional encoding is added to the input before it is passed into the transformer model, because otherwise the attention mechanism would be order invariant. However, both t