This paper explores large language model (LLM)-based intelligent agents, focusing on their definitions, methods, and future prospects. LLM-based agents, which use natural language as an interface, demonstrate strong generalization capabilities across various applications, from autonomous task assistance to coding, social, and economic domains. The paper provides an in-depth overview of LLM-based agents in single-agent and multi-agent systems, covering their definitions, research frameworks, and foundational components such as composition, cognitive and planning methods, tool utilization, and responses to environmental feedback. It also discusses mechanisms for deploying LLM-based agents in multi-agent systems, including multi-role collaboration, message passing, and strategies to alleviate communication issues between agents. The paper also highlights popular datasets and application scenarios, and concludes by envisioning future prospects for LLM-based agents in the evolving landscape of AI and natural language processing. LLM-based agents are defined as entities capable of perceiving their environment and taking actions. They can be categorized into five types: Simple Reflex agents, Model-based Reflex agents, Goal-based agents, Utility-based agents, and Learning agents. Reinforcement Learning (RL)-based agents and LLM-based agents fall under the category of Learning agents. LLM-based agents have several advantages over other agents, including potent natural language processing and comprehensive knowledge, zero-shot or few-shot learning, and organic human-computer interaction. LLM-based agents can be classified into single-agent and multi-agent systems. Single-agent systems are LLM-based intelligent agents proficient in handling multiple tasks and domains. Multi-agent systems (MAS) are computerized systems composed of multiple interacting intelligent agents. The paper discusses the characteristics of single-agent and multi-agent systems, including their application domains, memory and reconsideration mechanisms, data prerequisites, modalities, and toolsets. The paper also discusses the performance evaluation of LLM-based agents, including datasets and benchmarking methods. It examines the application of LLM-based agents across various domains, including natural sciences, social sciences, engineering systems, and general domains. The paper concludes by discussing the developmental trajectories of agents, including enhancing the adaptive capacity of LLM-based agents, incorporating multimodal models or large multimodal models (LMMs) to endow agents with multimodal information processing capabilities, and addressing the challenges encountered.This paper explores large language model (LLM)-based intelligent agents, focusing on their definitions, methods, and future prospects. LLM-based agents, which use natural language as an interface, demonstrate strong generalization capabilities across various applications, from autonomous task assistance to coding, social, and economic domains. The paper provides an in-depth overview of LLM-based agents in single-agent and multi-agent systems, covering their definitions, research frameworks, and foundational components such as composition, cognitive and planning methods, tool utilization, and responses to environmental feedback. It also discusses mechanisms for deploying LLM-based agents in multi-agent systems, including multi-role collaboration, message passing, and strategies to alleviate communication issues between agents. The paper also highlights popular datasets and application scenarios, and concludes by envisioning future prospects for LLM-based agents in the evolving landscape of AI and natural language processing. LLM-based agents are defined as entities capable of perceiving their environment and taking actions. They can be categorized into five types: Simple Reflex agents, Model-based Reflex agents, Goal-based agents, Utility-based agents, and Learning agents. Reinforcement Learning (RL)-based agents and LLM-based agents fall under the category of Learning agents. LLM-based agents have several advantages over other agents, including potent natural language processing and comprehensive knowledge, zero-shot or few-shot learning, and organic human-computer interaction. LLM-based agents can be classified into single-agent and multi-agent systems. Single-agent systems are LLM-based intelligent agents proficient in handling multiple tasks and domains. Multi-agent systems (MAS) are computerized systems composed of multiple interacting intelligent agents. The paper discusses the characteristics of single-agent and multi-agent systems, including their application domains, memory and reconsideration mechanisms, data prerequisites, modalities, and toolsets. The paper also discusses the performance evaluation of LLM-based agents, including datasets and benchmarking methods. It examines the application of LLM-based agents across various domains, including natural sciences, social sciences, engineering systems, and general domains. The paper concludes by discussing the developmental trajectories of agents, including enhancing the adaptive capacity of LLM-based agents, incorporating multimodal models or large multimodal models (LMMs) to endow agents with multimodal information processing capabilities, and addressing the challenges encountered.