The rapid advancement of large language models (LLMs) is transforming higher education by reshaping how learners engage with information, knowledge creation, and learning processes. This chapter provides a multidisciplinary cognitive perspective on the contrasting roles LLMs can play in structured and exploratory learning. We conceptualize two hypothetical use-cases situated at opposite ends of the task-agency continuum. The first, agentic guided instruction, characterizes LLMs as proactive tutors that provide tailored scaffolding during structured learning tasks. The second, epistemic information foraging, positions LLMs as low-agency tools that assist learners in gathering, interpreting, and synthesizing diverse ideas during open-ended exploratory projects. Drawing on cognitive load theory, we argue that the effectiveness of LLM-based tutoring depends on the interplay between task type (structured vs. exploratory), learner expertise (novice vs. advanced), and learning phase (early-stage vs. later-stage). For novice learners with limited domain knowledge, LLMs may act as knowledgeable interlocutors supporting lower-order educational goals, such as remembering, understanding, and applying. For advanced learners, however, the same degree of scaffolding can become redundant or even counterproductive, increasing extraneous load and reducing autonomy and transfer (i.e., expertise reversal effect). We propose educational design principles to guide the application and integration of both the agentic guided instruction role as well as the epistemic forage role. Between these two ways that LLMs can be used lies a collaborative mode of human–AI interaction, in which learners and LLMs co-construct understanding through dialogue, reflection, and shared agency—paving the way toward more productive forms of human–AI co-creation in higher education.