Having watched some of the neural network descriptions the game developer has posted, I believe it could easily be emergent if the simulation runs though enough generations.
The neural network can be pre-programmed (and the starting settings offer some basic neural pre-wiring), but has enough sensor inputs and behavior outputs that "If pushing a food pellet, swim to the right at an angle" is a possible behavior. This results in a clockwise swim direction that mechanically nudges other pellets towards the center of the circle.
Other inputs like "swim towards creatures of color #0000FF in your vision cone" (assuming the creature is blue) and "swim towards food in your vision cone" could have the second-order effect of "gathering food towards creatures of the same color -- which are presumably of similar species"
The neural network can be pre-programmed (and the starting settings offer some basic neural pre-wiring), but has enough sensor inputs and behavior outputs that "If pushing a food pellet, swim to the right at an angle" is a possible behavior. This results in a clockwise swim direction that mechanically nudges other pellets towards the center of the circle.
Other inputs like "swim towards creatures of color #0000FF in your vision cone" (assuming the creature is blue) and "swim towards food in your vision cone" could have the second-order effect of "gathering food towards creatures of the same color -- which are presumably of similar species"