In this study, we investigated human causal learning in a continuous time and space setting. We find participants to be capable active causal structure learners, and with the help of computational modelling explore how they mitigate the complexity of continuous dynamics data to achieve this. We propose participants combine systematic interventions with a narrowed focus on causal dynamics that occur during and directly downstream of their interventions. %identifying direct outgoing links from the variable they intervene on, attending only to the dynamics that occur during their interventions. This task-decomposition approach achieves comparable accuracy to attending to all the dynamics, while discarding almost half of the data. We argue this strategy makes sense from a resource rationality perspective: ignoring dynamics outside of interventions saves computational cost while the interventions naturally decompose the global learning problem into a series of more manageable sub-problems. We also find that when the causal relata are given real world labels, participants will use their domain specific priors to guide their structure inferences. In particular, individuals with accurate prior expectations were less likely to make the common local computations error of mistaking an indirect for a direct relationship. Overall, our experiments reinforce the idea that humans are frugal and intuitive active learners who combine actions and inference to optimize learning while minimizing effort.