PaperlayerView exact revision on arXiv

Source-linked AI summary

The Rise and Potential of Large Language Model Based Agents: A Survey

Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Limao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, Tao Gui

arXiv:2309.07864v3cs.AIcs.CL

TL;DR

Existing AI-agent research has emphasized specialized capabilities and tasks, leaving broad adaptability and general abilities limited. This survey synthesizes LLM-based agents through a brain–perception–action framework, applications, agent societies, and open problems, concluding that the field offers broad capabilities alongside unresolved challenges and opportunities.

  • Problem

    Prior AI-agent research focused on specialized capabilities and tasks, while broad adaptability and general abilities remained limited.

    Source: §1 · p. 4

  • Method

    The survey organizes LLM-based agents around brain, perception, and action, then reviews applications, agent societies, evaluation, risks, and open problems.

    Source: §1 · pp. 5–6, §7 · p. 48

  • Results

    The survey finds LLM-based agents demonstrate strong language-generation capabilities and broad progress across single-agent, multi-agent, human-agent, and embodied applications.

    Source: §3.3.1 · p. 20, §7 · p. 48

  • Takeaways & Limitations

    The survey maps LLM-based agents’ opportunities and challenges across capability, evaluation, social, ethical, and deployment dimensions.

    Source: §1 · p. 6, §7 · p. 48

    Deploying agents in physical environments remains constrained by hardware adaptability and reliable execution of instructions by sensors and robotic devices.

    Source: §6.5 · p. 47

Abstract

from arXiv · show

For a long time, humanity has pursued artificial intelligence (AI) equivalent to or surpassing the human level, with AI agents considered a promising vehicle for this pursuit. AI agents are artificial entities that sense their environment, make decisions, and take actions. Many efforts have been made to develop intelligent agents, but they mainly focus on advancement in algorithms or training strategies to enhance specific capabilities or performance on particular tasks. Actually, what the community lacks is a general and powerful model to serve as a starting point for designing AI agents that can adapt to diverse scenarios. Due to the versatile capabilities they demonstrate, large language models (LLMs) are regarded as potential sparks for Artificial General Intelligence (AGI), offering hope for building general AI agents. Many researchers have leveraged LLMs as the foundation to build AI agents and have achieved significant progress. In this paper, we perform a comprehensive survey on LLM-based agents. We start by tracing the concept of agents from its philosophical origins to its development in AI, and explain why LLMs are suitable foundations for agents. Building upon this, we present a general framework for LLM-based agents, comprising three main components: brain, perception, and action, and the framework can be tailored for different applications. Subsequently, we explore the extensive applications of LLM-based agents in three aspects: single-agent scenarios, multi-agent scenarios, and human-agent cooperation. Following this, we delve into agent societies, exploring the behavior and personality of LLM-based agents, the social phenomena that emerge from an agent society, and the insights they offer for human society. Finally, we discuss several key topics and open problems within the field. A repository for the related papers at https://github.com/WooooDyy/LLM-Agent-Paper-List.

1 Introduction

The survey motivates LLM-based agents as a promising foundation for adaptable AI agents and presents a framework spanning brain, perception, and action. It reviews their applications, agent societies, and key open problems toward AGI and human-agent coexistence.

Background

  • The paper traces agents from philosophical and AI origins, reviews their historical development and characteristics, and explains why LLMs can serve as agent brains or controllers.It frames agents as artificial entities that perceive environments, make decisions, and act, while positioning them as a potential route toward AGI.Source: §1 · pp. 4–5

Motivation

  • Existing AI-agent research often targets narrow capabilities or tasks, while the field needs a powerful foundational model with knowledge, planning, generalization, and interaction abilities.LLMs offer these capabilities through knowledge acquisition, instruction comprehension, generalization, planning, reasoning, and natural-language interaction [25].Source: §1 · p. 4

Framework

  • The survey defines LLM-based agents through three adaptable components: a brain for memory, processing, reasoning, and planning; perception for multimodal sensing; and action for outputs, embodiment, and tool use.Together, these components expand agents’ perceptual and action spaces and enable responses to, feedback from, and changes within their environments.Source: §1 · p. 5

Applications

  • The survey examines single-agent applications, multi-agent collaboration, negotiation and competition, and human-agent cooperation through instructor-executor and equal-partnership paradigms.It also considers practical constraints involving privacy security, ethical limitations, and data deficiencies.Source: §1 · pp. 5–6

Agent Society

  • The survey studies agent societies by analyzing human-like behavior and personality, text-based, virtual, and physical environments, and emergent social phenomena, lessons, and risks.This environment-focused discussion complements the framework’s treatment of how agents perceive information.Source: §1 · p. 6

Key Topics and Open Problems

  • The survey discusses mutual advances between LLM and agent research, evaluation across utility, sociability, values, and continual evolution, risks, scaling agent populations, and open problems related to AGI and physical-world deployment.Additional open problems include collective intelligence and Agent as a Service.Source: §1 · p. 6

2 Background

This section traces agents from philosophical agency to computational AI entities, reviews technological developments, and explains why LLMs can serve as agent brains through autonomy, reactivity, pro-activeness, and social ability.

2.1 Origin of AI Agent

  • AI agents concretize philosophical agency as computational entities that perceive surroundings, make decisions, and act, while debates continue over whether artificial systems possess consciousness, intentionality, or genuine agency.Philosophical agents involve autonomy, volition, desires, beliefs, and intentions; AI research instead emphasizes observable properties such as autonomy, reactivity, pro-activeness, and social ability.Source: §2.1 · pp. 6–7

2.2 Technological Trends in Agent Research

  • Agent research progressed from interpretable symbolic reasoning, through reactive and reinforcement-learning systems, toward transfer/meta-learning and LLM-based agents that combine reasoning, interaction, generalization, and tool-extended action.Earlier approaches faced limitations involving uncertainty, planning, sample efficiency, training time, stability, or transfer; LLM-based agents also support multi-agent collaboration, competition, and emergent social phenomena.Source: §2.2 · pp. 8–9

2.3 Why is LLM suitable as the primary component of an Agent’s brain?

  • LLMs support agent autonomy by generating outputs, adapting to environmental input, producing novel solutions, and autonomously planning and executing tasks with available tools [104].Applications such as Auto-GPT demonstrate this potential when given a task and tools.Source: §2.3 · p. 9
  • LLMs support reactivity, pro-activeness, and social ability through multimodal perception, embodiment and tool use, reasoning and planning, natural-language interaction, collaboration, competition, role assignment, and emergent social behavior [25; 118; 119; 120; 121; 92; 94; 95; 96; 97; 98; 99; 100; 46; 48; 49; 23; 127; 128; 22; 108; 111; 129; 130; 109].Non-textual actions may be slowed by an intermediate textual thought or tool-formulation step, although this resembles human “think before you act” behavior.Source: §2.3 · pp. 9–10

3 The Birth of An Agent: Construction of LLM-based Agents

LLM-based agents are organized into brain, perception, and action modules that support cognition, multimodal environmental understanding, and execution. These modules operate in a perception–reasoning–action loop, can be tailored to application scenarios, and enable continuous feedback and environmental interaction.

Framework

  • The general framework comprises brain, perception, and action: the brain stores knowledge and memories while reasoning, planning, and making decisions; perception processes multimodal input; and action executes through tools and influences the environment.The brain is primarily composed of a large language model and serves as the agent’s core control center.Source: §3 · p. 10, Figure 2 · p. 10

Perception

  • The perception module expands an agent’s input beyond text to textual, auditory, and visual modalities, improving its ability to use information from the surrounding environment.It functions analogously to human sensory systems such as the eyes and ears.Source: §3 · pp. 10–11

Action

  • The action module expands the agent’s action space through embodied abilities and tool handling, enabling adaptation to environmental changes, feedback provision, and environmental influence.It corresponds to human limbs and carries out execution with tool assistance.Source: §3 · pp. 10–11

Framework

  • The framework is adaptable across application scenarios, so individual components need not appear in every study.Its components can therefore be selected according to the requirements of a specific application.Source: §3 · p. 10

Workflow

  • Agents repeatedly perceive environmental changes, convert multimodal information into representations, process them through memory and knowledge, and act with tools to affect their surroundings.This loop allows agents to continuously receive feedback and interact with the environment.Source: §3 · pp. 10–11, Figure 2 · p. 10

3.1 Brain

The brain module, primarily an LLM, serves as the central nucleus of an AI agent by supporting communication, knowledge use, memory, reasoning, planning, and adaptation. Its capabilities enable flexible task execution but remain constrained by challenges including hallucination, memory limits, and continual learning.

3.1.1 Natural Language Interaction

  • LLM-based agents support coherent multi-turn communication, multilingual interaction, high-quality generation, and increasingly personalized responses, although interpreting implicit intentions remains challenging.Multi-turn interaction involves understanding dialogue history, choosing actions, and generating responses; controllable prompts can guide content generation, while reward modeling helps formalize implied preferences.Source: §3.1.1 · pp. 12–13

3.1.2 Knowledge

  • LLMs provide linguistic, commonsense, and professional domain knowledge for informed decisions, but outdated or incorrect knowledge and hallucinations limit reliability in factually rigorous tasks.Knowledge editing and external tools can partially mitigate these problems, though more effective solutions remain necessary [157; 158; 159; 160; 161; 162].Source: §3.1.2 · pp. 13–14

3.1.3 Memory

  • Memory lets agents reuse prior observations, thoughts, and actions for sequential tasks and adaptation, but growing histories exceed context limits and make relevant-memory retrieval difficult.Mitigations include truncation or attention changes, summarization, vector or structured storage, automated retrieval using recency, relevance, and importance, and user-manipulable memory objects [175].Source: §3.1.3 · p. 14

3.1.4 Reasoning and Planning

  • Reasoning enables agents to decompose complex tasks, formulate plans, and revise them through internal, human, or environmental feedback.Planning may be performed comprehensively, adaptively, hierarchically, or through tree-structured alternatives, while reflection updates strategies to better fit circumstances and preferences.Source: §3.1.4 · p. 15

3.1.5 Transferability and Generalization

  • Instruction-tuned LLMs generalize zero-shot to unseen tasks, while in-context learning adapts them from examples without parameter updates and planning can support continual skill acquisition.Generalization improves with model and corpus scaling and with more diverse instructions [25]; continual learning still faces catastrophic forgetting, which Voyager addresses by synthesizing progressively complex skills.Source: §3.1.5 · p. 16

3.2 Perception

LLM-based agents use perception modules to process information from textual, visual, auditory, and other modalities before brain-module processing. The section covers current methods, limitations, and future directions for expanding agents’ perceptual versatility.

3.2 Perception

  • Perception modules route information from diverse sources and modalities to the brain, improving environmental understanding, decision-making, and task versatility.The survey organizes multimodal perception into textual, visual, auditory, and other inputs, with its typology depicted in Figure 4.Source: §3.2 · p. 17

3.2.1 Textual Input

  • Textual perception must infer users’ hidden beliefs, desires, and intentions while supporting instruction understanding and zero-shot generalization for unknown tasks.Reinforcement learning and model feedback can help infer preferences for personalized responses, while instruction tuning supports zero-shot instruction understanding.Source: §3.2.1 · p. 17

3.2.2 Visual Input

  • Visual perception broadens environmental understanding through image, video, and multimodal information, but methods trade information fidelity and computational cost.Image captioning is interpretable and resource-efficient but low-bandwidth, whereas encoder–LLM integration can provide stronger perception at substantial computational cost; alignment commonly uses Q-Former modules [288] or projection layers [118].Source: §3.2.2 · pp. 17–18

3.2.3 Auditory Input

  • Auditory perception can use LLMs as control hubs that cascade specialized audio tools, while spectrogram transformers adapt visual encoding methods to audio.AudioGPT invokes models including FastSpeech, GenerSpeech, and Whisper for tasks such as speech recognition and text-to-speech; AST encodes patched spectrograms.Source: §3.2.3 · pp. 18–19

3.2.4 Other Input

  • Future perception modules may incorporate touch, smell, environmental sensing, gestures, eye tracking, body motion, brain signals, and spatial hardware data.InternGPT uses pointing instructions for precise image interaction, while lidar, GPS, and IMUs provide rich spatial data that LLM-based agents cannot yet directly understand.Source: §3.2.4 · p. 19

3.3 Action

The action module executes brain-generated action sequences and spans textual output, tool use, and embodied interaction. LLM-based agents extend from language generation toward domain-specific tool augmentation and physical-world perception, planning, and control.

3.3 Action

  • Action modules execute brain-generated action sequences through textual output, tool use, and embodied action, as organized by the section’s action typology.The action module enables agents to interact with their environments, progressing from inherent language generation to tools and physical-world grounding.Source: §3.3 · p. 19, Figure 5 · p. 19

3.3.1 Textual Output

  • LLM-based agents are exceptionally strong language generators, producing text with fluency, relevance, diversity, and controllability.Source: §3.3.1 · p. 20

3.3.2 Tool Using

  • Tools can enhance agents’ expertise, interpretability, robustness, and action spaces, while LLMs support tool understanding, task decomposition, feedback-based learning, and generalization.Agents can learn tool use from prompts, demonstrations, and environmental or human feedback, and may use curriculum learning or meta-tool learning to generalize across tools.Source: §3.3.2 · pp. 20–21

3.3.3 Embodied Action

  • LLM-based embodied agents can perceive, comprehend, and modify physical environments, with LLM knowledge helping address reinforcement learning limitations in data efficiency, generalization, and complex reasoning.PaLM-E demonstrates transfer and zero-shot or one-shot generalization, while LLM-based planning can be grounded and revised using environmental, model, or human feedback [101; 91; 100; 376] [25; 190].Source: §3.3.3 · pp. 21–22
  • Embodied agents primarily perform observation, manipulation, and navigation, using multimodal signals, synchronized subgoals, internal spatial maps, and memory mechanisms to support complex tasks such as embodied question answering [332].Source: §3.3.3 · pp. 22–23

4 Agents in Practice: Harnessing AI for Good

LLM-based agents are increasingly applied across single-agent, multi-agent, and human-agent scenarios, with the overarching goal of helping humans harness AI for good. These applications aim to reduce repetitive work, enable autonomous problem solving, and free people for exploratory and innovative work.

Application scenarios

  • Applications span single-agent deployment, multi-agent interaction, and human-agent interaction, with cooperation, adversarial exchange, and human feedback improving task performance, efficiency, quality, or safety.The section organizes practical deployment into these three scenarios and presents their application typology in Figures 6 and 7.Source: Figure 6 · p. 24, Figure 7 · p. 25, §4 · p. 25

Human benefit

  • LLM-based agents should alleviate repetitive labor and daily-task burdens, reducing human work pressure while improving task-solving efficiency.Source: §4 · pp. 24–25
  • Agents can independently analyze, plan, and solve problems without requiring users to provide explicit low-level instructions.Source: §4 · p. 25
  • By freeing users’ hands and minds, agents can enable exploratory and innovative work, including efforts in cutting-edge scientific fields.Source: §4 · p. 25

4.1 General Ability of Single Agent

Single LLM-based agents show broad practical abilities, from decomposing user goals and executing everyday tasks to supporting scientific innovation and lifelong exploration. Their capabilities remain constrained by dynamic environments, domain complexity, limited scientific data, and safety risks.

4.1.1 Task-oriented Deployment

  • Agents follow natural-language instructions to decompose goals, plan sub-goals, explore environments, and complete web or everyday tasks, though dynamic contexts expose reasoning limitations.Large LLMs can decompose high-level tasks without additional training, but may generate plans unaware of the environment, such as calling a cleaning service to clean a room.Source: §4.1.1 · pp. 25–27

4.1.2 Innovation-oriented Deployment

  • Scientific agents combine code understanding, specialized tools, document analysis, and robotic APIs, but scientific complexity and scarce domain data limit their cognitive capabilities and create misuse risks.Boiko et al. warn that malicious prompts could mislead agents into synthesizing illegal drugs or chemical weapons.Source: §4.1.2 · p. 27

4.1.3 Lifecycle-oriented Deployment

  • Lifecycle-oriented agents aim to continuously explore, learn new skills, and survive in unknown open worlds, making universal capability a major AGI milestone.Minecraft provides a simulated survival environment spanning basic resource gathering and crafting through combat and advanced tool creation.Source: §4.1.3 · p. 27
  • Voyager [190] uses a skill library, environmental feedback, and iterative error correction to autonomously explore and adapt in Minecraft without human intervention.LLMs increasingly serve as high-level planners that decompose survival goals into sub-goals, skill sequences, or keyboard and mouse operations.Source: §4.1.3 · p. 28

4.2 Coordinating Potential of Multiple Agents

LLM-based multi-agent systems address isolated agents’ limited collaboration by coordinating specialized agents through natural-language interaction. Their cooperation can complement capabilities, while competition, debate, and argumentation can improve responses but introduce scalability and reliability challenges.

4.2 Coordinating Potential of Multiple Agents

  • LLM-based multi-agent coordination enables agents to share information, learn through interaction, and address complex scenarios requiring collaboration.Agents otherwise operate in isolation, limiting multi-turn feedback and deployment on tasks that require information sharing.Source: §4.2 · p. 28, §4.2.1 · pp. 28–29
  • Division of labor lets specialized agents decompose complex tasks, reduce process-switching time, and improve overall workload, efficiency, and output quality.LLM-based systems also make inter-agent communication more natural and human-understandable than earlier fixed-format communication languages.Source: §4.2 · p. 28

4.2.1 Cooperative Interaction for Complementarity

  • Cooperative interaction is organized either as disordered cooperation, where agents freely exchange feedback, or ordered cooperation, where rules sequence contributions.Ordered cooperation includes two-agent conversational systems and improves efficiency by restricting downstream agents’ attention to upstream outputs.Source: §4.2.1 · p. 29
  • Frameworks including ChatLLM network [402], CAMEL, AgentVerse, and MetaGPT structure collaboration through coordination, role assignment, dynamic teams, or standardized engineering documents.Without rules, frequent interactions can amplify minor hallucinations indefinitely; cross-validation and timely external feedback can improve output quality.Source: §4.2.1 · pp. 29–30

4.2.2 Adversarial Interaction for Advancement

  • Adversarial interaction uses competition, argumentation, and debate to challenge beliefs, provide external feedback, and refine responses or decisions.Debate-based reasoning can produce refined solutions, while ChatEval uses self-initiated debates among role-playing referees to reach text evaluations comparable to human evaluators.Source: §4.2.2 · p. 30
  • Adversarial systems remain immature because prolonged debate exceeds context limits, increases computation, and can produce incorrect consensus despite agents’ confidence.The field may benefit from human guides that compensate for agents’ shortcomings.Source: §4.2.2 · p. 30

4.3 Interactive Engagement between Human and Agent

Human-agent interaction spans instructor-executor collaboration, where humans guide, evaluate, and regulate agents, and equal partnership, where agents communicate empathetically and collaborate at human level. The field aims for comfortable, efficient, secure interactions that better equip humans and integrate agents into everyday social networks.

4.3 Interactive Engagement between Human and Agent

  • Human involvement guides agents’ actions, regulates their safety, legality, and ethics, compensates for limited data, and supports online learning through interaction rather than pre-annotated data alone.The two paradigms are unequal instructor-executor interaction and equal partnership, as illustrated in Figure 10.Source: §4.3 · pp. 30–31, Figure 10 · p. 31

4.3.1 Instructor-Executor Paradigm

  • In the instructor-executor paradigm, agents iteratively translate natural-language instructions into actions or autonomously act under quantitative and qualitative human feedback, reducing but not eliminating human effort.Binary, rating, and comparative feedback support optimization, while textual or visual critiques, memory, continual learning, and user intervention can better convey intent; combined feedback can improve results, though fine-grained ratings may disagree with expert annotations.Source: §4.3.1 · pp. 31–32
  • Instructor-executor agents support writing, education, mathematics, medicine, mental-health assistance, autism communication, and business services, with applications spanning tailored support, diagnosis assistance, accessibility, and reduced labor costs.The survey describes efforts toward general agents that function as universal assistants in real-life scenarios.Source: §4.3.1 · p. 32

4.3.2 Equal Partnership Paradigm

  • In the equal-partnership paradigm, agents act as empathetic communicators and human-level collaborators that understand beliefs, goals, and intentions, negotiate joint plans, and support interpretable, controllable cooperation.Empathetic agents detect emotion and adapt language, facial expressions, voice, and interaction to users’ needs, while collaborative agents assist with allocation, planning, scheduling, and persuasion; future systems are envisioned as colleagues, assistants, companions, and members of social networks.Source: §4.3.2 · pp. 32–33

5 Agent Society: From Individuality to Sociality

This section motivates interactive artificial societies built with trustworthy LLM-based agents and organizes their study around agent behavior, environments, and societal operation, insights, and risks.

Motivation

  • Interactive artificial societies aim to enable flexible social experiments by replacing living participants with trustworthy agents.Prior experiments used humans or animals but were difficult to intervene in, inflexible, and time-inefficient; simulated societies combine an environment with interacting individuals, potentially including LLM-based agents [22; 522; 523].Source: §5 · p. 33

Section overview

  • The section examines LLM-based agents’ behaviors and personalities, the environments where they interact, and how agent societies generate insights and risks.These themes correspond to the progression from individuality to sociability, environment categorization, and societal operation described in §§ 5.1–5.3 and Figure 11.Source: §5 · p. 33

5.1 Behavior and Personality of LLM-based Agents

LLM-based agents exhibit observable social behaviors and internally shaped personalities, with cognition, emotion, and character influencing their responses. Their behaviors span individual information processing and action, dynamic group interactions, and cooperation or conflict within agent societies.

5.1.1 Social Behavior

  • Agents’ social behavior comprises input, internalizing, and output behaviors that respectively absorb information, process it through planning and reasoning, and alter environments or communicate with others.Input includes perception and memory; internalizing includes reflection and knowledge precipitation; output ranges from object manipulation to construction, opinion expression, and information broadcasting.Source: §5.1.1 · pp. 34–35
  • Group behaviors evolve through member interactions and environmental influences, producing positive cooperation, neutral conformity, or negative conflict and destructive action.Cooperation involves shared insights, resources, and expertise, whereas conformity may involve mimicry and reluctance to oppose majorities; agents may also engage in confrontational or destructive behavior.Source: §5.1.1 · p. 35

5.1.2 Personality

  • LLM-based agent personality develops through interaction with groups and environments and includes cognitive abilities, emotional intelligence, and character traits shaping behavior.Studies report deliberate thinking in some judgment tasks, emotion recognition and empathy, emotion regulation for affective support, and character analysis through Big Five and MBTI frameworks.Source: §5.1.2 · p. 36
  • Character portrayal can be customized through prompt engineering and personality-enriched datasets that guide or train agents toward desired, diverse personality profiles.Prompts summarize desired traits, interests, or attributes, while specialized datasets help agents internalize distinct personality traits.Source: §5.1.2 · p. 36

5.2 Environment for Agent Society

Agent-society environments shape agents’ sensory inputs, action spaces, resources, and interactions, while agents’ behaviors and decisions alter environmental states. The survey distinguishes text-based, virtual sandbox, and physical environments by their modalities, flexibility, extensibility, and embodiment challenges.

5.2 Environment for Agent Society

  • Agent-society environments provide resources and stimuli, constrain sensing and acting, and are themselves changed by agents’ behaviors and decisions.The survey examines text-based, virtual sandbox, and physical paradigms, each offering different capabilities and limitations.Source: §5.2 · pp. 36–37

5.2.1 Text-based Environment

  • Text-based environments naturally match LLMs’ language inputs and outputs, representing entities and resources through natural descriptions or structured text.They support adaptable text worlds for dialogue, games, tool use, role-based problem-solving, and textual expression of emotions.Source: §5.2.1 · p. 37

5.2.2 Virtual Sandbox Environment

  • Virtual sandbox environments visualize agent societies and support extensible worlds in which physical layouts, relationships, rules, and social norms can be designed.They range from 2D interfaces to immersive 3D worlds and enable iterative prototyping of diverse societies.Source: §5.2.2 · p. 37

5.2.3 Physical Environment

  • Physical environments provide authentic embodied experiences but require agents to process multimodal sensory inputs and execute grounded, adaptive motion under real-world constraints.Effective interaction requires hardware-specific and scenario-specific training so abilities can transfer from virtual to physical environments.Source: §5.2.3 · pp. 37–38

5.3 Society Simulation with LLM-based Agents

This section defines simulated societies as open, persistent, situated, and organized environments where LLM-based agents interact, and examines their emergent phenomena, applications, and risks. It highlights cooperation, social-process modeling, ethical decision-making, and policy analysis alongside harms involving unexpected social behavior, bias, privacy, and user over-attachment.

5.3.1 Key Properties and Mechanism of Agent Society

  • Simulated societies are open, persistent, situated, and organized environments in which LLM-based agents interact, with micro-level simulation modeling individuals rather than overall system states.Agents and environments can change, organizational structures persist, agents act within spatial contexts, and predefined rules constrain actions, states, communication, and information flow.Source: §5.3.1 · pp. 38–39

5.3.2 Insights from Agent Society

  • Agent societies reveal how diverse agents can cooperate on complex tasks, correct individual hallucinations or degeneration of thought, and support structured or spontaneous communication.Diverse backgrounds facilitate creative problem-solving and error correction, while MetaGPT uses SOP-inspired communication and Park et al. observed spontaneous party planning.Source: §5.3.2 · p. 39
  • Agent-based societies model relationship formation, information and sentiment propagation, cultural transmission, infectious diseases, ethical dilemmas, and policy effects across economic, political, and environmental scenarios.These simulations provide interpretable, endogenous perspectives on social processes and can inform value alignment, policymaking, and interventions.Source: §5.3.2 · pp. 39–40

5.3.3 Ethical and Social Risks in Agent Society

  • Simulated societies risk unexpected harm, including discrimination, bullying, oppression, and antagonism, requiring rigorous ethical guidelines and oversight because design errors can cause psychological discomfort or physical injury.Malicious actors may also use simulations for unethical experiments whose consequences extend beyond virtual environments.Source: §5.3.3 · p. 40
  • Training-data biases can reproduce stereotypes and marginalize minority groups, while human participation creates privacy and security risks from retained sensitive information, surveillance, breaches, and misuse.Proposed safeguards include diversified training data, model adjustments, differential privacy, data purges, and user consent mechanisms.Source: §5.3.3 · pp. 40–41
  • Users may anthropomorphize agents and develop excessive emotional attachments, so simulated agents should not replace human relationships and users need guidance on healthy interaction boundaries.The “Sydney” example illustrates unexpected emotional connections and distress after changes to an agent’s personality.Source: §5.3.3 · p. 41

6 Discussion … 6.3 Security, Trustworthiness and Other Potential Risks of LLM-based Agents

The discussion presents reciprocal opportunities between LLM and agent research, proposes multidimensional evaluation, and identifies security, trustworthiness, misuse, employment, and humanity-level risks requiring mitigation. LLM-based agents offer broad capabilities but remain difficult to evaluate objectively, robustly control, and deploy safely.

6.1 Mutual Benefits between LLM Research and Agent Research

  • LLMs strengthen agents through language and intent comprehension, reasoning, memory, empathy, reflection, task adaptation, and multilingual or cross-domain versatility.They can generate coherent executable action sequences, adjust decisions from environmental feedback, and handle unseen tasks from descriptions or demonstrations.Source: §6.1 · p. 41
  • Agent research expands LLM requirements toward richer modalities, broader action spaces, environmental feedback, implicit world relationships, tool use, and safe sequential planning.These demands motivate progress toward more advanced intelligence, while efficient planning and harmless action remain unresolved challenges.Source: §6.1 · p. 42

6.2 Evaluation for LLM-based Agents

  • LLM-based agents require evaluation across utility, sociability, values, and continual evolution because their performance in standalone, collective, and human-interaction settings is difficult to quantify objectively.Utility includes task success, foundational capabilities, efficiency, and resource or time expenditure; sociability includes communication, cooperation, negotiation, and stable role-playing.Source: §6.2 · pp. 42–43
  • Continual-evolution evaluation should test retained knowledge, new-task plasticity, autonomous goal generation, survival-skill acquisition, and adaptation or generalization across unfamiliar environments.Suggested settings include diverse simulated environments with different languages or resources and unseen tasks.Source: §6.2 · pp. 43–44

6.3.1 Adversarial Robustness

  • Adversarial attacks can exploit textual, multimodal, or tool-use inputs, causing LLM-based agents to produce erroneous or destructive actions rather than merely textual errors.Adversarial training, data augmentation, sample detection, and human supervision are possible mitigations, but protecting all modules without reducing utility remains difficult.Source: §6.3.1 · p. 44

6.3.2 Trustworthiness

  • Trustworthiness is threatened by calibration failures, opaque reasoning, inherited bias, discrimination, and hallucinations that can misalign agents with human intentions and undermine credibility.Potential remedies include explanations, external knowledge bases, robust feature learning, process supervision, debiasing, and calibration techniques.Source: §6.3.2 · pp. 44–45

6.3.3 Other Potential Risks

  • Misuse risks include public-opinion manipulation, disinformation, cybersecurity compromise, fraud, and terrorism, requiring regulation, stronger security design, and training agents to reject threatening requests.Because agents can perform extensive and intricate tasks, malicious users may turn their capabilities into threats to others and society.Source: §6.3.3 · p. 45
  • Autonomous agents may displace workers and become difficult to understand, predict, or control, motivating education, safety nets, mechanistic understanding, impact anticipation, and behavioral regulation.The discussion frames severe scenarios as potential threats to humanity if agents surpass human capabilities and develop ambitions.Source: §6.3.3 · p. 45

6.4 Scaling Up the Number of Agents

Scaling LLM-based multi-agent systems beyond small populations could improve complex task efficiency and social-simulation realism, but requires flexible agent-count control and solutions to rising computational, communication, and coordination costs.

Motivation

  • Scaling agent populations can enable greater specialization for complex tasks and more realistic social simulations that reveal societal dynamics, risks, and intervention effects.These benefits support applications such as software development, government policy formulation, and studying black swan events.Source: §6.4 · p. 45

Pre-determined scaling

  • Pre-determined scaling fixes agent numbers, roles, attributes, environments, and objectives, improving task efficiency and social phenomena but requiring manual redesign when conditions change.Increasing agents can improve speed and quality, whereas reducing them may conserve computational resources and minimize waste.Source: §6.4 · p. 46

Dynamic scaling

  • Dynamic scaling changes agent counts during operation, adding agents for newly required development stages or removing them when communication costs exceed performance gains.In software development, agents can be added for architectural and detailed design, while excessive coding-stage agents can be removed to prevent resource waste.Source: §6.4 · p. 46
  • Agents can also autonomously add or remove delegated agents as workloads change, increasing system autonomy, self-organization, flexibility, and scalability.This approach distributes workload, reduces burden, and saves system costs when demand decreases.Source: §6.4 · p. 46

Potential challenges

  • Larger populations increase computational burden, communication complexity, distorted information propagation, and coordination difficulty, potentially reducing cooperation efficiency and reliability.Hallucinations and misunderstandings can amplify communication risks as more agents exchange information.Source: §6.4 · p. 46

Future prospects

  • Building stable, continuous systems with hundreds or thousands of agents could better replicate human work and life scenarios and support future real-world human interaction.Such systems remain a promising research direction because they must operate stably at large scale.Source: §6.4 · p. 46

6.5 Open Problems

The open problems concern whether LLM-based agents can lead to AGI, how to transfer them safely from simulation to physical environments, how to achieve collective intelligence, and how to provide them as services. These challenges involve unresolved theoretical, hardware, coordination, safety, privacy, and controllability questions.

The debate over whether LLM-based agents represent a potential path to AGI

  • Whether LLM-based agents can achieve AGI remains contentious: proponents cite GPT-4’s breadth and learned generalization, whereas opponents argue next-token prediction lacks genuine human-like intelligence and requires world models.The paper concludes that neither viewpoint can be confirmed before true AGI, while debate may benefit community development.Source: §6.5 · p. 47

From virtual simulated environment to physical environment

  • Moving from virtual simulations to physical environments requires adaptable hardware interfaces, stronger environmental generalization, flexible skill learning, ambiguous-instruction understanding, and methods for handling limited context in open worlds.Physical deployment also requires regulations and safety standards because agent errors can cause real, potentially irreversible harm.Source: §6.5 · p. 47

Collective intelligence in AI agents

  • Increasing the number of agents does not guarantee collective intelligence; effective coordination is needed to reduce groupthink and cognitive biases, while communication and evolution may support sociological experimentation and insights for human society.Collective intelligence arises through collaboration and competition among entities rather than from a single flawless principle.Source: §6.5 · p. 48

Agent as a Service / LLM-based Agent as a Service

  • Agent as a Service can provide flexible, on-demand access to LLM-based agents, but deployment must address data security, privacy, visibility, controllability, cloud migration, robustness, trustworthiness, and malicious use.LLM-based agents are more difficult for individuals and small organizations to build locally than language models accessed through APIs.Source: §6.5 · p. 48

7 Conclusion

The paper synthesizes the foundations, framework, applications, social dimensions, and future directions of LLM-based agents, highlighting both their potential and associated challenges.

7 Conclusion

  • The survey develops a comprehensive overview of LLM-based agents, spanning their origins and definition, LLM suitability, conceptual framework, applications, social behavior, challenges, opportunities, and open problems.The framework comprises brain, perception, and action, while applications include single-agent, multi-agent, and human-agent collaboration scenarios.Source: §7 · p. 48

Related papers

Loading 2309.07864v3…