An LLM's core function is to predict the next token given preceding text, repeating this process to generate responses. All higher-level features such as chat interfaces, coding assistants, and tool‑calling agents are built as scaffolding around this token‑prediction loop.