Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects

1.1 Intelligent Agents

• Agents autonomously perceive environments, decide, and act; LLM-based and reinforcement-learning agents are learning agents that improve behavior through experience [5].Learning from environmental observations and action outcomes can address limited adaptability and multi-step decision-making challenges.Source: §1.1 · p. 4

1.2 RL-based Agents

• RL agents learn policies through trial and error to maximize cumulative rewards, but face slow training, high sample demands, instability, and weak generalization to new tasks [7, 12, 13].Transfer learning may help, but effective techniques for RL remain an open research challenge.Source: §1.2 · pp. 4–5

1.3 LLM-based Agents

• LLMs face context-length constraints, delayed knowledge updates, and inability to directly use external tools; combining their language comprehension and generation with agents’ decision-making and planning offers promising resolutions to practical-application obstacles.Source: §1.3 · pp. 5–6
• The survey covers LLM-based agent systems and frameworks, datasets and evaluation, applications across four domains, and future directions including adaptation and multimodal capabilities.It also identifies challenges encountered by agents as a future research concern.Source: §1.3 · p. 6

2.1 Single-Agent System

• A single-agent system uses an LLM-based agent for multiple tasks and domains, represented as V = (L, O, M, A, R), with memory, actions, and reflection guiding later decisions.The LLM plans and decides from observations, memory, and rewards; actions interact with tools or the environment, whose feedback is recorded for the agent’s rethink process.Source: §2.1 · pp. 6–7

2.2 Multi-Agent System

• MAS frameworks characterize coordination through agent granularity, knowledge heterogeneity, control distribution, and communication protocols, while application taxonomies separate system- and agent-level characteristics.Decker outlines four dimensions; Parunak distinguishes system function, system architecture, and agent architecture.Source: §2.2 · p. 7
• MAS classifications also distinguish heterogeneity and communication, while MARL taxonomies cover task mode, agent type, learning style, and knowledge sharing [28].The MARL learning styles include independent learning, CTDE, and fully centralized learning.Source: §2.2 · p. 7
• LLM-based multi-agent systems are modeled as graphs of agents linked by message passing, and categorized by multi-role coordination and planning type.The proposed coordination categories are cooperative, competitive, mixed, and hierarchical; planning types are CPDE and DPDE.Source: §2.2 · p. 8

2.3 Agent System Template

• Agent templates support development through data, training, and evaluation workflows, customizable planning and review functions, and configurable roles for multi-agent systems.Examples include ToolLLM [74] and open-source projects such as AutoGPT, and multi-agent templates such as AutoGen [83].Source: §2.3 · p. 8

3.1.1 Planning

• Planning methods guide LLMs through complex tasks using prompted reasoning, external planners, or staged processes.In-context methods include decomposing tasks into steps, exploring alternative reasoning paths, and refining outputs; external methods connect LLMs to symbolic planners, while multi-stage methods separate rapid and deliberative planning.Source: §3.1.1 · p. 12

3.1.2 Memory

• Memory systems combine short-term context, long-term stores, and retrieval methods to preserve and access information for task execution.Long-term memory can use external knowledge sources, skill repositories, or accumulated experience; retrieval and memory selection can be adapted to the agent’s tasks, and both memory categories may be used concurrently.Source: §3.1.2 · pp. 13–14

3.1.3 Rethinking

• Rethinking methods use environmental feedback, self-learning, and reflection to improve agents’ reflection and re-planning capabilities.Approaches include in-context interaction, supervised feedback, reinforcement learning, and coordination among modules.Source: §3.1.3 · pp. 14–15

3.1.4 Environments

• Agents interact through computer, gaming, coding, real-world, and simulation environments, learning and applying knowledge through natural language and environmental feedback.Simulation environments offer controlled but realistic settings for experimentation and solution development, with potential transfer of knowledge and skills to real-world applications.Source: §3.1.4 · p. 18

3.1.5 Action

• Action capabilities let agents communicate with environments through text generation or use tools such as APIs, calculators, code interpreters, and physical actions.Tool-use approaches connect LLMs to external tools for task planning and execution, including routing natural-language requests and invoking APIs.Source: §3.1.5 · p. 18
• Chameleon [156] selects and combines inventory modules through a natural-language planner, while AVIS formulates tool strategies and inspects outputs to gather knowledge for answers.Source: §3.1.5 · p. 18
• Tool-planning frameworks support multi-turn tool-aided reasoning (ChatCoT), structured execution (TPTU [162]), DFS-based API-path search (ToolLLM [74]), and configurable model and plugin integration (Gentopia [163]).Source: §3.1.5 · p. 18
• Tool-creation frameworks generate task-appropriate tools through staged generation and execution (Cai et al. [164]) or retrieve and specialize general-purpose toolkits for complex tasks (CRAFT).Source: §3.1.5 · p. 18

3.2.1 Relationship of Multi-Agent Systems

• LLM-based MAS research examines cooperative and competitive dynamics, mixed cooperation and competition, and hierarchical task decomposition through parent and child agents.Cooperative approaches emphasize role and task allocation for collaborative decisions; hierarchical research remains developing, with only a limited number of levels explored.Source: §3.2.1 · pp. 19–21

3.2.2 Planning Type

• CPDE centralizes planning before agents execute independently, whereas DPDE gives each agent local planning and coordination through communication or negotiation.CPDE can optimize globally but risks complexity, single-point failure, and slow adaptation; DPDE offers robustness and adaptability but may sacrifice global optimality and incur coordination overhead.Source: §3.2.2 · pp. 21–22
• DPDE agents can exchange information without communication, through explicit messages, or via shared memory, each with distinct coordination and overhead tradeoffs.No communication minimizes overhead but can impair collaboration; explicit communication supports coordination but adds delays and bandwidth costs; shared memory streamlines exchange but can create contention and scalability challenges.Source: §3.2.2 · p. 22
• Shared memory provides a unified information mechanism, including a central knowledge base or shared model parameters, but parameter sharing can cause overspecialization.MetaGPT [67] uses a global memory pool that agents can query or update; dynamically adjusting shared weights can balance specialization and generalization.Source: §3.2.2 · p. 22

3.2.3 Methods of Enhancing Communication Efficiency

• Communication efficiency can be improved with structured protocols and mediator models that selectively limit costly or redundant agent interactions.Protocols can clarify message structure, with embeddings or JSON as possible formats; mediator decisions can depend on task complexity, inter-agent relationships, and communication costs.Source: §3.2.3 · p. 23
• Strategies for inaccurate LLM outputs include auxiliary fine-tuning with synthetic data and Chain of Verification, which checks a draft before producing a refined response.The section presents these approaches as ways to mitigate flattering or unfounded outputs and hallucinations in LLM-based MAS.Source: §3.2.3 · pp. 23–24

4.1 Dataset

• Because most LLM-based agents need no further LLM training and some task-specific datasets are inaccessible, the survey lists publicly available, extensively used datasets.Source: §4.1 · p. 24

4.2 Benchmark

• No widely used benchmark for LLM-based agents exists, though studies compare agents and researchers are developing potential future evaluation standards.Source: §4.2 · p. 24
• ToolBench [74] covers single- and multi-tool utilization; TE [199] assesses human-behavior emulation, while Akata et al. [200] study LLM social behavior.Source: §4.2 · p. 24
• Ziems et al. [201] provide prompting best practices and a pipeline evaluating 13 language models zero-shot across 24 representative CSS benchmarks.Source: §4.2 · p. 24
• AgentSims [143] supports open-source LLM evaluation; Drori et al. assess mathematical reasoning, BMTools tool use, and SmartPlay game-based agent performance.Source: §4.2 · p. 24
• MLAgentBench benchmarks AI research tasks, MetaTool evaluates conscious tool choice, and LLM-Co assesses cooperative reasoning and long-term collaboration.Source: §4.2 · p. 24

• Figure 12 is titled “The Prospect of LLM-based agents.”Source: Figure 12 · p. 25

5.1.1 Mathematics

• In mathematics, LLM-agent work centers on reasoning and theoretical derivation, with tool-integrated problem solving and formal proof search among its approaches [40].Multi-agent systems may partition complex problems to improve computational speed and precision, but accuracy, knowledge representation, interpretability, and validation remain concerns.Source: §5.1.1 · pp. 25–26

5.1.2 Chemistry and Materials

• Chemistry and materials agents support real-world experiment automation, MOF prediction and generation, and chemical tasks, with broader potential in simulation, reaction optimization, and material design [39] [38].Coscientist uses search, code execution, and experiment automation; ChatMOF [39] supports retrieval, prediction, and structure generation; ChemCrow [38] accesses chemistry databases. Improving prediction and generation accuracy and reliability remains a challenge.Source: §5.1.2 · p. 26

5.1.3 Biology

• Biology research remains limited: BioPlanner [223] evaluates protocol generation and planning, while OceanGPT uses multi-agent collaboration to generate ocean-science data.Potential applications include ecosystem, group-behavior, and cellular or molecular simulations; agents must handle complex, dynamic biological systems and large, heterogeneous, noisy datasets.Source: §5.1.3 · p. 27

5.1.4 Climate Science

• Climate applications include extracting emissions data from ClimateWatch4, with proposed sensor-based forecasting and agent-based atmospheric simulation and optimization.Complex multi-agent simulations pose computational-efficiency challenges, while better planning and reconsideration and stronger numerical understanding could improve simulations and forecasts.Source: §5.1.4 · p. 27

5.2.1 General Task Assistant

• General task assistants range from multimodal, multi-task agents such as Generalist Agent [18] to frameworks including HuggingGPT [155], AutoGPT, and ModelScope-Agent [228].Other examples support natural-language software development, plugins and web scenarios, objective-driven task creation, multi-agent simulations, and chat-based function invocation.Source: §5.2.1 · p. 28
• Key challenges include efficient task-decomposition granularity, memory use and information integration, adaptable tools, and long-term learning in changing tasks and environments.Future work may improve automated task decomposition and optimization, and integrate images, sounds, and video into task assistants.Source: §5.2.1 · p. 28

5.2.2 Work/Research Assistant

• Work/research assistants can browse webpages, databases, and literature, summarize materials, generate experimental code for validation, and draft conclusions.The workflow may support reports, papers, presentations, and narrative or cinematic works.Source: §5.2.2 · p. 28
• Examples span multi-agent debate for text generation in ChatEval [16], autonomous research reporting in GPT Researcher, and scientific experiment design and execution by Boiko et al. [65].Other systems generate episodic content, layouts, and music workflows, or use summary trees for long-text queries in MemWalker.Source: §5.2.2 · p. 28
• Compared with general task assistants, work/research assistants need stronger memory and knowledge integration, domain-specific tools for validation, and cross-domain knowledge and skills.The section also identifies originality and creativity, including avoiding repetitive or overly similar content, as challenges.Source: §5.2.2 · pp. 28–29
• Future work/research assistance may extend artistic creation in music and film and use human-machine collaboration to draw on human knowledge for more original work.The stated aim is to provide greater convenience for human work and creativity.Source: §5.2.2 · p. 29

5.3.1 Economics and Finance

• LLM-based agents are being used to study economic behavior, principal-agent conflict, auctions, and financial trading, with market simulation and financial analysis identified as further opportunities.Economic modeling also faces challenges in representing actors’ states and interaction spaces, validating anthropomorphic behavior, and scaling agent populations; reinforcement learning may reduce LLM interactions.Source: §5.3.1 · pp. 29–30

5.3.2 Politics

• Political research uses LLM-based agents to analyze politicians’ speeches, detect ideology, predict voting, and explore strategic trade-offs, while proposed uses include political forecasting and policy analysis.Key challenges include modeling political uncertainty, ensuring efficient and effective communication, and keeping agent behavior ethically sound.Source: §5.3.2 · p. 30

5.3.3 Society

• LLM-based agents can simulate credible human behavior and social interactions, participate in group chats, learn in simulated settings, and model opinion dynamics.Challenges include coherent turn-taking, socially aligned training, and diverse, personalized simulations of human actors.Source: §5.3.3 · p. 31

5.3.4 Law

• Legal-domain agent research remains limited, with work on simulating Supreme Court decisions and retrieval-supported case learning; proposed uses include legal assistance and decision analysis.These applications require longer context and stronger memory, legal-knowledge representation, and accurate legal reasoning.Source: §5.3.4 · pp. 31–32

5.3.5 Psychology

• Psychology applications include mental-health support and experiment simulation, with studies reporting on-demand, nonjudgmental support as well as distortions and self-interest tendencies in simulated behavior.Challenges include harmful-content filtering, consistent and anthropomorphic communication, and user overdependence.Source: §5.3.5 · p. 32

5.3.6 Education

• Education research uses LLM-based agents for mathematical research assistance and classroom simulation, while future policy simulations could assess academic input, quality, and equity.Challenges include harmless and credible outputs, personalized teaching, and better understanding of students’ questions, expressions, and emotions.Source: §5.3.6 · pp. 32–33

5.3.7 Management

• Management applications include simulating company roles and assessing agents’ workflow comprehension, collaborator selection, and task delegation; proposed areas include organizational work and supply-chain optimization.Challenges include handling multiple levels, roles, objectives, scaling, and varied historical and real-time data.Source: §5.3.7 · p. 33

5.4.1 Computer System

• Computer-system agents span natural-language computer control, human interaction, coding and testing, security, recommendation, and games, but need better code quality, tool use, and adaptation to changing problems.Source: §5.4.1 · pp. 33–34

5.4.2 Robotics System

• Robotics research centers on task planning and control using language feedback, perception, world models, and teammate prediction; multi-robot and UAV applications also require stronger adaptation and real-time multimodal response.Source: §5.4.2 · pp. 34–35

5.4.3 Power System

• LLM-based agents are a nascent energy-research area, with proposed uses in smart-grid management, distributed-resource scheduling, and energy markets, while integration must handle diverse energy types, grid levels, and interoperability.Source: §5.4.3 · pp. 35–36

5.4.4 Transportation System

• Transportation research is nascent, using LLM reasoning for traffic dynamics, traffic-data decision support, and autonomous driving; signal control and vehicle simulation remain constrained by real-time needs and multi-objective decisions.Source: §5.4.4 · p. 36

5.4.5 Industrial Control System

• Industrial applications combine LLMs with digital twins and automation for production planning and control, alongside lighting and chip-design uses; deployment faces text-to-real-world planning gaps and scalability across roles and objectives.Source: §5.4.5 · pp. 36–37

5.4.6 Medical System

• Medical-agent research remains scarce, covering epidemic simulation, hallucination mitigation, and clinical-task evaluation, with prospective uses in epidemiology and drug discovery constrained by biological complexity and accuracy.Source: §5.4.6 · p. 37

5.4.7 Military System

• Military-agent research is limited, with prospective uses in war simulation, tactical analysis, intelligence, and decision support requiring realistic environments, human collaboration, and attention to legal and ethical issues.Source: §5.4.7 · pp. 37–38

6.1 Trend

• Objective evaluation remains challenging, motivating standardized foundational and domain-specific benchmarks for LLM-based agents.Assessments should cover planning, decision-making, tool and retrieval use, and memory; domain benchmarks should test agents in execution environments and specialized fields, with datasets and human feedback for psychological outcomes.Source: §6.1 · p. 38
• Agents need continual learning, self-evaluation, dynamic goal-setting, and adaptability to changing tasks, environments, and user preferences.The discussion calls for learning from text, user interactions, and real-world experience while retaining prior knowledge; agents should use environmental feedback to identify weaknesses, track improvement goals, and adjust their behavior and communication styles.Source: §6.1 · pp. 38–39
• Multimodal agents may move from converting images into text toward LMMs that plan and reconsider directly from image inputs.Earlier frameworks commonly route multimodal inputs through text and use multimodal models for output; direct image-based planning could improve information-use efficiency and multimodal task processing.Source: §6.1 · p. 39

6.2.1 Intrinsic Constraints of LLMs

• LLM context-length limits and hallucinated, invalid, or erroneous outputs can undermine agent and multi-agent-system performance.Long articles or complex dialogues may cause essential information to be lost, while fluent but incorrect or irrelevant generations can impair overall efficacy.Source: §6.2.1 · p. 39

6.2.2 Dynamic Scaling

• LLM-based multi-agent systems need to scale dynamically across hardware and software environments while adjusting to demand.Scaling requires adapting agent counts, memory sizes, and computing and storage resources; for systems where one LLM allocates roles and plans actions, its reasoning costs must also be managed, including through LLM cascades.Source: §6.2.2 · pp. 39–40

6.2.3 Security and Trust

• Appropriate permissions and security are critical because excessive authority can cause harmful actions and erode human and enterprise trust.Permission mechanisms should support collaboration without exceeding agents’ authority, and reliability testing such as ToolEmu can detect failures and quantify risks across tools and scenarios.Source: §6.2.3 · p. 40

7 Conclusion

• The review traces the path from agents through RL-based agents to LLM-based agents, covering definitions, planning, memory, rethinking, action, and external environments.It also surveys multi-agent systems’ role relationships, planning types, and enhanced communication methods.Source: §7 · p. 40
• The paper discusses development prospects and challenges across fields, including LLM limitations, the dynamic expansion of multi-agent systems, and security and trust issues.It proposes possible solutions to these challenges.Source: §7 · p. 40
• The authors conclude that LLM-based agents may be a significant step forward, although current research remains far from achieving AGI.Source: §7 · p. 40