Human behaviour is extraordinarily flexible. Task fMRI and patient studies highlight a network of frontoparietal brain regions, called “multiple-demand” regions, that serve as a hub for flexible cognition. Neurons within this network adapt to code what is relevant for the current task, across many task types. Under the attentional episodes view, prioritising immediately relevant information in this way allows us to break complex tasks into simple parts and drive neural resources toward the problem at hand. Many studies demonstrate that relevant information is preferentially encoded, such that we can read out features more accurately when they are task-relevant. Yet we do not know whether the preferential coding that we see on slow timescales and in simple tasks supports moments of narrow focus, or “temporal modules”, in more complex, multi-part tasks. This is central to the attentional episodes account: that selection of immediately relevant information, through preferential coding, dynamically shifts to give us the information that we need for each part of a task. Chapter 2 begins by asking how preferential coding emerges in a multi-part task. Using MEG data from a dual-epoch task with single object (Experiment 1) and dual-object (Experiment 2) displays, I show that what is relevant can be preferentially encoded in sequential task epochs with similar rapidity. Preferential coding in either epoch was only detected with dual-object displays, mirroring dominant theories of attention as a spatial spotlight, or a filter to reduce complexity. Chapter 3 builds on this by tracing how ventral visual and multiple-demand regions contribute to, and communicate, coding of relevant stimulus information throughout a task. I resolve MEG data from a multi-epoch visual attention task (Experiment 2) to source space, to pull apart how the stimulus coding in Chapter 2 arises in visual and domain-general regions. Using Granger causality, I probe the timecourse of top-down and bottom-up information flow as the relevant feature shifts. I show feedback from prefrontal to visual regions emerging in both task epochs, again highlighting how flexibly we are able to direct focus to multiple parts of a task in turn. Chapter 4 extends Chapters 2 and 3 to a situation like those we face often in daily life, where what is relevant for each part of a task can be present throughout. These distractors with some task-relevance are also common in classical tests of fluid ability, and could be preferentially attended if selection is not strictly directed to the immediate task part. I use a behavioural task with two sequential displays, each containing a relevant- and an irrelevant-coloured moving dot cloud. Despite being cued to attend to two colours in sequence, participants were not more distracted by the second target colour when it appeared as a distractor in the first task epoch. That is, we can effectively direct our focus to what is immediately relevant, even when presented with a future-relevant feature. Attending to what is currently relevant as we move through parts of a task is a central aspect of flexible behaviour. These studies probe the limits of this temporal modularity in attention. They show that we can preferentially encode distinct stimulus features as what is relevant changes; and that we can overtly respond to what is relevant in each task part, even when a feature relevant for one task part is visible throughout. Together, they emphasise our extraordinary capacity to direct our focus toward what is relevant.