You can use simulators. However the problem is that if you're for example running material science experiments, those simulations will consume a lot of compute and take weeks, so spamming different approaches in the way an agent tends to work might not work quite as well.