Characterizing the cognitive architecture of human mindreading forces us to address two puzzles in people's attributions of belief: Why children show inconsistent expectations about others' belief‐based actions, and why adults' reasoning about belief is sometimes automatic and sometimes not. The seemingly puzzling data suggest that humans have many mindreading systems that use different models of mental representations. The efficient system is shared by infants, children, and adults, and uses a minimal model of the mind, which enables belief‐like states to be tracked. The flexible system develops late and uses a canonical model, which incorporates propositional attitudes. A given model's operation has signature limits that produce performance contrasts, in children as well as adults, between certain types of mindreading tasks.