The policy network is the agent, the part that looks at a position and picks a card. What makes Hokm a genuine test is what the network is not allowed to see. It gets an encoding of the observable state only: its own hand, the trump suit, and the cards played so far. The other three hands, its partner's included, stay hidden.
That is the hard part of the game and the reason it is worth learning at all. When you cannot see the full state, the best move depends on belief about the unseen part, on what the play so far implies your opponents are holding, not just on the cards in front of you. The network has to fold that inference into a single decision from a fixed-width view of the visible world.
A policy that only reacted to its own hand would be playing a different, easier game. This one has to decide under partial observability, which is the same thing that makes real decisions hard: acting well on what you can see while reasoning about what you cannot.