Generative artificial intelligence (AI) is entering oncology. Large language models are the near-term workhorse because oncology runs on narrative text and structured tables. We review current adoption and outline a practical path from stand-alone chat models to retrieval-augmented systems and, ultimately, agentic assistants that plan tasks, call domain tools, and integrate multimodal data within the electronic health record. Concrete uses include molecular tumor board synthesis with transparent evidence, grading along guidelines, synoptic radiology and pathology drafting, and computable trial matching. We also map the constraints: fragmented hospital information technology, privacy and provenance requirements, domain shift across sites, and persistent hallucinations. We envision that evaluation must move beyond leaderboards toward multicenter, prospective designs with endpoints that reflect clinical utility, such as faithfulness to cited sources, extraction accuracy, time to task completion, plan correctness, recovery after tool failure, and silent clinical studies before exposure. Finally, we sketch an adoption trajectory. Institutions will replace ad hoc use of public tools with sanctioned drafting assistants, then embed retrieval and calculators inside the record, and only later enable event-driven agents that propose context-aware actions. The destination is augmentation, not automation: a learning assistant that shows its work, improves routine care, and leaves clinical judgment with clinicians.