agent who only "knows" about the world. For this agent, every perception is represented as a change in the model of the world

As far as I see, for such an agent there still would be a (likely implicit) computation that converts the input signals into updates on the world model; which, pretty much, would be the equivalent of the bridge hypothesis. In addition, there has to be just as implicit information regarding “where in the world model the agent's inputs and outputs are located” (to compute the outputs if nothing else).

As a result, it would be equivalent to an agent with a built-in bridge hypothesis.