Trying it in your head first
The real payoff of a world model is planning. If an agent can predict what a candidate sequence of actions will lead to, it can evaluate several possible plans internally and pick the one that looks best, all without touching the real environment. This is often described as the agent "imagining" possible futures: it rolls its world model forward in simulation, action by action, and checks where each rollout ends up.
This is a meaningful shift from an agent that only learns by direct trial and error. Real-world trial and error can be slow, expensive, or outright dangerous, think of a robot arm that has to physically attempt a risky grasp to learn it fails. An agent with a decent world model can rule out many bad plans in imagination before ever attempting one for real.