A striking limitation of human cognition is our inability to execute some tasks simultaneously. Recent work suggests that such limitations can arise from a fundamental trade-off in network architectures that is driven by the sharing of representations between tasks: sharing promotes quicker learning, at the expense of interference while multitasking. From this perspective, multitasking failures might reflect a preference for learning efficiency over parallel processing capability. We explore this hypothesis by formulating an ideal Bayesian agent that maximizes expected reward by learning either shared or separate representations for a task set. We investigate the agent's behavior and show that over a large space of parameters the agent sacrifices long-run optimality (higher multitasking capacity) for short-term reward (faster learning). Furthermore, we construct a general mathematical framework in which rational choices between learning speed and processing efficiency can be examined for a variety of different task environments.