An LLM is not an AI agent. An LLM predicts useful next tokens. An agent is the system around it: state, retrieval, tools, permissions, retries, and observability. Treat the model as a decision engine—not your whole application.