LLM-based Agents replace the traditional Policy Network with a pre-trained language model, lowering the barrier to Agent development from requiring specialized reinforcement learning engineers to developers who can write prompts.