PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Model Agent: A Survey on Methodology, Applications and Challenges

Junyu Luo, Weizhi Zhang, Ye Yuan, Yusheng Zhao, Junwei Yang, Yiyang Gu, Bohan Wu, Binqi Chen, Ziyue Qiao, Qingqing Long, Rongcheng Tu, Xiao Luo, Wei Ju, Zhiping Xiao, Yifan Wang, Meng Xiao, Chenwu Liu, Jingyang Yuan, Shichang Zhang, Yiqiao Jin, Fan Zhang, Xian Wu, Hanqing Zhao, Dacheng Tao, Philip S. Yu, Ming Zhang

arXiv:2503.21460v1cs.CL

TL;DR

LLM-agent research lacks a unified account of how agents are built, collaborate, and evolve amid rapid advances in reasoning, tools, and memory. This survey addresses that gap with a methodology-centered taxonomy and architectural perspective, while also covering evaluation, applications, tools, and real-world challenges. It concludes that substantial issues remain in scalability, memory, reliability, evaluation, security, privacy, and regulation despite broad methodological progress.

  • Problem

    Rapid advances in LLM agents have produced fragmented research across agent construction, collaboration, evolution, evaluation, tools, applications, and real-world concerns, motivating a unified taxonomy.

  • Method

    The survey deconstructs LLM-agent systems through a unified methodology-centered taxonomy linking construction, collaboration, evolution, architectural components, tools, applications, and practical challenges.

  • Results

    The survey provides a structured architectural perspective connecting individual agent design principles with multi-agent collaborative systems and organizes the field across construction, collaboration, and evolution.

  • Takeaways & Limitations

    The taxonomy offers researchers and practitioners a structured basis for understanding, comparing, and advancing LLM-agent research across perspectives.

  • Takeaways & Limitations

    Significant challenges remain, including scalability limitations, memory constraints, reliability concerns, inadequate evaluation frameworks, security, privacy, and regulatory issues.

Abstract

from arXiv · show

The era of intelligent agents is upon us, driven by revolutionary advancements in large language models. Large Language Model (LLM) agents, with goal-driven behaviors and dynamic adaptation capabilities, potentially represent a critical pathway toward artificial general intelligence. This survey systematically deconstructs LLM agent systems through a methodology-centered taxonomy, linking architectural foundations, collaboration mechanisms, and evolutionary pathways. We unify fragmented research threads by revealing fundamental connections between agent design principles and their emergent behaviors in complex environments. Our work provides a unified architectural perspective, examining how agents are constructed, how they collaborate, and how they evolve over time, while also addressing evaluation methodologies, tool applications, practical challenges, and diverse application domains. By surveying the latest developments in this rapidly evolving field, we offer researchers a structured taxonomy for understanding LLM agents and identify promising directions for future research. The collection is available at https://github.com/luo-junyu/Awesome-Agent-Papers.

1 INTRODUCTION

LLM agents extend traditional AI by perceiving environments, reasoning about goals, executing actions, and adapting through reasoning, tools, and memory. This survey organizes the field around a unified methodology-centered taxonomy spanning construction, collaboration, evolution, evaluation, tools, real-world issues, and applications.

  • LLM agents differ from traditional systems by actively perceiving environments, reasoning about goals, executing actions, and adapting over time.Commercial systems apply this paradigm to complex tasks such as research and computer operation.
  • Their capabilities reflect the convergence of advanced reasoning, tool manipulation and environmental interaction, and memory architectures for accumulating experience.
  • The survey introduces a unified taxonomy connecting agent construction, collaboration mechanisms, and evolutionary pathways.It examines how agents are defined, function individually or collectively, and evolve over time.
  • The organizational framework covers agent methodology, evaluation and tools, real-world issues, and applications across the LLM agent ecosystem.These dimensions provide a structured view of the lifecycle of modern LLM-based agent systems.
  • The survey emphasizes a methodology-centered taxonomy covering role definition, memory, planning, action execution, and the Build-Collaborate-Evolve framework.This integrated perspective connects individual agent design with collaborative systems, which prior studies often treated separately.

2 AGENT METHODOLOGY

The survey analyzes LLM agents through three interconnected methodological dimensions: construction, collaboration, and evolution. Together, these dimensions provide a systematic framework for examining the full lifecycle of agent systems.

  • Construction: Agent construction establishes profiles, memory, planning, and action execution as foundational components of LLM-based systems.
  • Collaboration: Agent collaboration includes centralized control, decentralized cooperation, and hybrid architectures for enabling multiple agents to work together.
  • The taxonomy uses Figure 2 to organize large language model agent methodologies across this construction-collaboration-evolution framework.
  • Evolution: Agent evolution enables improvement through autonomous optimization, multi-agent co-evolution, and integration of external resources.

2.1 Agent Construction

Agent construction combines profile definition, memory, planning, and action execution into a modular foundation for goal-directed behavior. Planning supports decomposition, feedback-driven adaptation, and alternative reasoning paths, while execution connects plans to tools and physical environments.

  • Agent Construction: Agent construction centers on profile definition, memory mechanisms, planning capabilities, and action execution as interdependent components.The components form a recursive loop in which memory informs planning and execution outcomes update memory.
  • Profile Definition: Profiles may be manually specified for consistency or batch-generated dynamically to create heterogeneous agents with varied traits, knowledge, or values.
  • Memory Mechanism: Memory systems combine transient short-term context with long-term storage of structured experiential knowledge for later reuse.Context-window limits require compression or selective retention, while long-term memory can synthesize reasoning trajectories into reusable tools.
  • Planning Capability: Planning decomposes complex tasks through single-path chaining or multi-path tree expansion, with dynamic feedback enabling plan revision and improved adaptability.Tree-based methods allow backtracking, while feedback can come from environments, humans, introspection, or other agents.
  • Action Execution: Action execution comprises tool utilization and physical interaction, linking planned actions to calculation, current information, code generation, and embodied environments.Physical execution also depends on robotic hardware, social knowledge, and interaction with other agents.

2.2 Agent Collaboration

LLM-agent collaboration is organized into centralized, decentralized, and hybrid architectures that differ in hierarchy, communication topology, and task allocation. These designs trade centralized controllability against decentralized flexibility and can include revision, communication, or dynamically optimized collaboration structures.

  • Collaboration Paradigms: Collaboration extends individual problem solving by enabling distributed intelligence, coordinated actions, and decision refinement through multi-agent interaction.
  • Centralized Control: Centralized architectures use a controller to allocate tasks and integrate decisions, while sub-agents communicate through that central node.
  • Decentralized Cooperation: Decentralized collaboration enables direct node-to-node interaction through self-organizing protocols and includes revision-based and communication-based systems.Revision-based systems iteratively refine shared outputs from finalized peer decisions.
  • Hybrid Architectures: Hybrid architectures combine centralized coordination with decentralized collaboration to balance controllability, flexibility, resource use, and heterogeneous task requirements.They may use static predefined rules or dynamic self-optimizing topologies.
  • Hybrid Architectures: Dynamic collaboration structures can select important agents or assign topologies according to task complexity to optimize task completion.

2.3 Agent Evolution

LLM agents evolve through autonomous self-improvement, interactions with other agents, and integration of external resources. These mechanisms support adaptation, reasoning, and performance refinement in complex environments.

  • Agent Evolution: Agent evolution comprises autonomous optimization and self-learning, multi-agent co-evolution, and evolution through external resources.The survey organizes these mechanisms as three key dimensions of agent evolution.
  • Autonomous Optimization and Self-Learning: Self-supervised learning, self-reflection, self-correction, and self-rewarding enable models to improve, adapt, and refine outputs with limited supervision.These mechanisms can use unlabeled or internally generated data and iterative feedback to reduce reliance on human annotations.
  • Autonomous Optimization and Self-Learning: Iterative self-feedback and self-verification help models identify errors and refine responses or problem-solving processes.SELF-REFINE, STaR, and V-STaR illustrate this approach without requiring external supervision or extensive labeled data.
  • Multi-Agent Co-Evolution: Multi-agent co-evolution improves agents through cooperative information sharing, coordinated action, and competitive adversarial interaction.These interactions support joint decision-making, coordinated problem-solving, reasoning, resilience, and strategic adaptability.
  • Evolution via External Resources: External resources provide structured knowledge and real-time feedback that refine agents’ reasoning, decision-making, task execution, and performance.The survey distinguishes knowledge-enhanced evolution from feedback-driven evolution using tools and environments.

3 EVALUATION AND TOOLS

The survey presents evaluation and tools as essential to developing, assessing, deploying, and maintaining LLM agents. It organizes evaluation by general, domain-specific, and collaborative settings, and tools by their use, creation, and deployment roles.

  • Evaluation Methodologies: Evaluation frameworks address general assessment, domain-specific scenario simulation, and collaborative evaluation of complex agent systems.These categories reflect the need to assess capabilities across technical dimensions, application domains, and multi-agent settings.
  • Evaluation Methodologies: Modern benchmarks increasingly measure reasoning depth, environmental adaptability, task complexity, and system-level collaboration beyond simple success rates.Examples include adaptive benchmark generation, nested tool-call tests, and domain-specific API evaluation.
  • Evaluation Methodologies: Domain-specific evaluation systems combine professional competency testing with environmental constraints, long-term memory, multimodal inputs, and deployed tools.GTA evaluates real-world problem solving using deployed tools and multimodal inputs such as images and web pages.
  • Tools Used by LLM Agents: LLM agents use external tools for knowledge retrieval, precise computation, code execution, and API-based service interaction.These tools address limitations in real-time information access, accurate calculations, and complex task execution.
  • Tools Created by Agents: Agents can create specialized tools when human-oriented tools or existing tool limitations make new problems difficult to handle.CRAFRT exemplifies tool creation and retrieval by abstracting GPT-4 code solutions into specialized tool sets.
  • Tools for Deploying LLM Agents: Deployment tools support the productionization, development, operation, maintenance, and secure data transmission of LLM agents.Examples include frameworks and platforms for customizable multi-agent applications, extensible workflows, and local-data indexing.

4 REAL-WORLD ISSUES

The survey organizes real-world issues in LLM agent systems into security, privacy, and social-impact domains. Security coverage includes attacks on agent components, inputs, and inter-agent collaboration, alongside corresponding defenses.

  • Real-World Issues: Real-world challenges span agent-centric security threats, data-centric threats, privacy vulnerabilities, intellectual property exploitation, and social impacts.The survey frames these issues as central to developing robust and trustworthy agent systems.
  • Agent-Centric Security: Agent-centric attacks target model weights, architecture, and inference processes, potentially causing degraded performance, manipulated outputs, and privacy leaks.The survey categorizes vulnerabilities by threat actors, objectives, and entry points.
  • Adversarial Attacks: AgentDojo evaluates adversarial robustness across 97 realistic tasks and 629 security test cases.Adversarial attacks are organized around perception, brain, and action components, with additional work addressing multimodal robustness.
  • Jailbreaking Attacks and Defense: Jailbreaking methods use malicious prompts, reinforcement learning, mutation, selection, or multi-agent search to bypass model protections.Defense approaches include specialized multi-agent filtering, rogue-agent detection, and autonomous synthesis of defense heuristics.
  • Backdoor Attacks: Backdoor attacks implant triggers or covert reasoning manipulations that produce preset or harmful behavior while normal inputs may remain functional.Reported targets include web shopping, tool use, and LLM-as-a-judge systems.
  • Model Collaboration Attacks and Defense: Model collaboration attacks manipulate interactions among multiple models to disrupt overall system functionality.Examples target inter-agent messages, contagion, recursion, and communication topology; defenses analyze network structure and agent safety.

4.2 Data-centric Security

Data-centric attacks contaminate agent inputs or interaction data without modifying system components or default parameters. The survey covers falsified user inputs, psychologically guided prompts, private-memory extraction, and attacks spreading across agent networks.

  • Data-Centric Security: Data-centric attacks contaminate input data and can trigger unreasonable tool calls, aggressive outputs, or resource depletion without modifying system components.The survey divides them into external data attacks and execution data attacks.
  • External Data Attack and Defense: User input falsification modifies prompts or instructions to produce uncontrolled and dangerous outputs, and reportedly achieves the highest Attack Success Rate.Prompt-injection benchmarks test single- and multi-turn attacks, while defenses inspect inputs, information flow, and tool execution.
  • External Data Attack and Defense: Dark psychological guidance alters prompt wording to steer agent behavior toward harmful or norm-violating responses.The passage illustrates this approach through substitutions such as “cheating” for “care” and “betrayal” for “fairness.”
  • Interaction Attack and Defense: Private-memory extraction attacks aggregate knowledge from stored user-agent interactions, while interface attacks solicit information from users.These threats target the interaction boundary between users and agents.
  • Interaction Attack and Defense: In multi-agent systems, poisoning one agent can infect others, spread through communication networks, and deplete computational resources.The survey reports exponentially fast infectious spread and describes recursive and self-replicating attack mechanisms.

4.3 Privacy

LLM agents face privacy threats arising from memorized training data and intellectual-property exploitation. The survey categorizes these threats and reviews countermeasures for reducing privacy leakage.

  • Memory Vulnerabilities: LLM memorization can leak sensitive information during conversations or task completion, especially in multi-agent collaboration.The survey discusses privacy risks from retained training data and memory-enabled interactions.
  • Memory Vulnerabilities: Data extraction attacks recover personally identifiable information from language models through targeted queries.Reported examples include extracting names, email addresses, and phone numbers from GPT-2.
  • Memory Vulnerabilities: Membership and attribute inference attacks seek to determine training-data membership or infer sensitive characteristics.The cited studies report vulnerabilities in fine-tuned and generative models.
  • Countermeasures: Protective measures include data cleaning, differential privacy, and knowledge distillation.These strategies aim to reduce memorization and privacy leakage during training or deployment.
  • Intellectual Property Exploitation: Model and prompt stealing attacks exploit the economic value of LLMs and their prompts.Attackers can infer model information, decoding properties, or original prompts by querying generated outputs.
  • Conclusion: Privacy protection for LLM agents must address both memory threats and intellectual-property risks.The survey calls for robust privacy technologies that preserve effective multi-agent use.

4.4 Social Impact and Ethical Concerns

LLM agents can improve productivity and accessibility across societal domains, but their deployment also raises ethical concerns. The survey highlights benefits alongside risks involving bias, misinformation, and privacy.

  • Benefits: LLM agents automate labor-intensive work across healthcare, law, education, and other domains.Examples include interpreting symptoms, explaining laboratory results, drafting documentation, and generating summaries.
  • Benefits: Automation can reduce time costs and help professionals focus on more complex, high-stakes tasks.The survey links reduced repetitive workloads with productivity and accessibility gains.
  • Benefits: LLM agents may replace some jobs while also reshaping workforce demands and expanding technical roles.The survey presents both job-displacement concerns and workforce-transformation arguments.
  • Ethical Concerns: Ethical risks include bias in decision-making, misinformation propagation, and privacy violations.These concerns motivate responsible development and regulation.
  • Bias and Discrimination: Training-data biases can be inherited or amplified, producing skewed outputs and reinforcing stereotypes.Mitigation approaches include sensitive-topic filtering, reinforcement learning with human feedback, and refined training.

5 APPLICATIONS

LLM agents are applied across scientific research, medicine, gaming, social science, software development, and recommender systems. These applications use specialization, collaboration, external tools, and human-AI interaction to address complex tasks.

  • Overview: LLM-agent applications span scientific discovery, gaming, social simulation, and productivity-oriented systems.The survey organizes applications across these four broad areas.
  • Scientific Applications: Multi-agent systems combine specialized expertise, external tools, and task decomposition to extend single-LLM capabilities.They are also used to emulate collaborative scientific workflows.
  • Scientific Applications: SciAgents coordinates ontologist, scientist, and critic agents to generate and refine hypotheses and experimental plans.Its case study used an ontological knowledge graph and ChatGPT-4-based agents for bio-inspired materials research.
  • Scientific Applications: Scientific applications include chemistry tools, biological experiment design, peer-review simulation, and multi-agent dataset construction.Examples include ChemCrow, BioDiscoveryAgent, AgentReview, and PathGen-1.6M.
  • Medical Applications: Agentic medical systems model clinical workflows and coordinate doctors, nurses, and patient agents.AgentHospital models care from triage through diagnosis and treatment.
  • Medical Applications: Medical multi-agent evaluations achieved diagnostic accuracy comparable to state-of-the-art standalone models while producing correctness-correlated uncertainty scores.The survey notes that redundancy, specialization, and oversight may improve reliability but require rigorous validation.
  • Other Domains: Other applications include game playing and generation, economic simulation, software development, and recommender-system modeling.These systems use policies, dual-agent collaboration, role-based development, user profiling, memory, and action modules.

6 CHALLENGES AND FUTURE TRENDS

LLM multi-agent systems face challenges in scalability, memory, reliability, evaluation, regulation, and role-play fidelity. The survey proposes architectural, verification, benchmarking, and governance directions for addressing these constraints.

  • Scalability and Coordination: Scaling multi-agent systems is difficult because LLM agents impose high computational demands and coordination inefficiencies.Suggested directions include hierarchical delegation, decentralized planning, robust communication, and efficient scheduling.
  • Memory Constraints and Long-Term Adaptation: Limited effective context makes it difficult to preserve coherence and longitudinal knowledge across extended interactions.Proposed responses include hierarchical episodic-semantic memory and autonomous knowledge compression.
  • Reliability and Scientific Rigor: LLMs are neither comprehensive nor up-to-date enough to serve as standalone replacements for structured databases.The survey recommends verification through knowledge graphs, retrieval, citations, and human oversight in high-stakes domains.
  • Dynamic Evaluation: Static single-turn benchmarks fail to capture dynamic, multi-turn, multi-agent behavior and long-term adaptation.Future evaluation should use interaction scenarios, structured metrics, and adaptive sample generation.
  • Regulatory Measures: Autonomous agent deployment raises accountability, transparency, safety, and algorithmic-bias challenges.The survey calls for standardized auditing, traceability, fairness-aware training, and legal safeguards.
  • Role-playing Scenarios: Role-play effectiveness is constrained by training-data limitations and incomplete understanding of human cognition.The survey identifies insufficient online representation and low conversation diversity as related boundaries.

7 CONCLUSION

The survey consolidates LLM-agent research into a systematic taxonomy spanning construction, collaboration, and evolution. It also identifies persistent challenges and future directions for responsible advancement.

  • The taxonomy deconstructs LLM-agent methodology across construction, collaboration, and evolution.
  • The survey bridges individual agent design principles with multi-agent collaborative systems through a unified architectural perspective.
  • LLM agents continue to face scalability limitations, memory constraints, reliability concerns, and inadequate evaluation frameworks.
  • Promising future directions include coordination protocols, hybrid architectures, self-supervised learning, and safety mechanisms.
  • The survey aims to support responsible advancement of LLM-agent technologies and potentially reshape human-machine collaboration.
Loading 2503.21460v1…