PaperlayerView exact revision on arXiv

Source-linked AI summary

From Language to Action: A Review of Large Language Models as Autonomous Agents and Tool Users

Sadia Sultana Chowa, Riasad Alvi, Subhey Sadi Rahman, Md Abdur Rahman, Mohaimenul Azam Khan Raiaan, Md Rafiqul Islam, Mukhtar Hussain, Sami Azam

arXiv:2508.17281v2cs.CL

TL;DR

LLM agents are developing rapidly, but evidence about their architectures, adaptation, capabilities, and evaluation remains fragmented. This review synthesizes recent research through seven questions, examining agent systems, tools, cognition, prompting, fine-tuning, benchmarks, and datasets. It concludes that external tools, dynamic evaluation, and attention to personalization, safety, and alignment are central to future development.

  • Problem

    Research on LLM agents lacks unified evidence across baseline models, architectures, tools, cognitive capabilities, adaptation methods, evaluation, and safety.

  • Method

    The review systematically synthesizes recent LLM-agent research across single- and multi-agent systems, external tools, reasoning, planning, memory, prompting, fine-tuning, benchmarks, and datasets.

  • Results

    External tool access is described as an essential foundation for LLM autonomy, while ReAct and Reflection dominate single-agent frameworks and evaluation is shifting toward dynamic, process-oriented methods.

  • Takeaways & Limitations

    Future LLM agents should address alignment, reliability, generalization, personalization, privacy, trust, and accountability alongside capability development.

  • Takeaways & Limitations

    Personalized agents introduce risks including anthropomorphism, over-trust, sensitive-information disclosure, and flawed attribution of accountability.

Abstract

from arXiv · show

The pursuit of human-level artificial intelligence (AI) has significantly advanced the development of autonomous agents and Large Language Models (LLMs). LLMs are now widely utilized as decision-making agents for their ability to interpret instructions, manage sequential tasks, and adapt through feedback. This review examines recent developments in employing LLMs as autonomous agents and tool users and comprises seven research questions. We only used the papers published between 2023 and 2025 in conferences of the A* and A rank and Q1 journals. A structured analysis of the LLM agents' architectural design principles, dividing their applications into single-agent and multi-agent systems, and strategies for integrating external tools is presented. In addition, the cognitive mechanisms of LLM, including reasoning, planning, and memory, and the impact of prompting methods and fine-tuning procedures on agent performance are also investigated. Furthermore, we evaluated current benchmarks and assessment protocols and have provided an analysis of 68 publicly available datasets to assess the performance of LLM-based agents in various tasks. In conducting this review, we have identified critical findings on verifiable reasoning of LLMs, the capacity for self-improvement, and the personalization of LLM-based agents. Finally, we have discussed ten future research directions to overcome these gaps.

1 Introduction

LLMs have evolved from language-processing tools into autonomous agents that plan, decide, act, and use external tools. This review organizes the field's architectures, capabilities, evaluation concerns, and research challenges.

  • Motivation: LLMs increasingly serve as autonomous agents that perform planning, decision making, and tool interaction in complex workflows.Few-shot, chain-of-thought, and self-ask prompting improved their potential beyond conventional NLP tasks.
  • Agentic systems: Single-agent systems use decision loops involving planning, memory, and tool use, but often struggle with dynamic environments and adaptive context tracking.These limitations motivate multi-agent approaches and external memory integration.
  • Organization: The paper is structured around a taxonomy spanning agent architectures, tools, cognitive capabilities, interaction paradigms, evaluation, safety, and human-agent dynamics.Figure 1 provides an overview of this taxonomy and the review's organization.
  • Review scope: The review examines architectures, external-tool interaction, single- and multi-agent frameworks, reasoning, planning, memory, prompting, fine-tuning, and evaluation.Its taxonomy covers the structure and interaction paradigms of LLM-based agentic systems.
  • Open challenges: The review identifies alignment, reliability, and generalization as fundamental challenges for improving the robustness and intelligence of LLM agents.It also organizes related concerns around safeguards, behavior control, failure modes, human collaboration, and social dynamics.

2 Related works

Existing surveys cover LLM-agent architectures and tool use but leave several dimensions fragmented. This review addresses those gaps through a unified seven-question comparison and taxonomy.

  • Existing surveys: Prior surveys classified LLM-agent reasoning, planning, tool use, modular architectures, memory, action control, and multi-agent communication.Their coverage established foundational taxonomies and reviewed agent frameworks and interaction strategies.
  • Research gaps: Existing surveys leave limited coverage of baseline LLM choice, prompting, fine-tuning, memory enhancement, and unified evaluation.The review identifies these as notable gaps in the literature.
  • Review framework: The review compares seven research questions covering baseline models, external tools, agent frameworks, cognitive capabilities, adaptation strategies, evaluation, and concerns.Table 1 structures the comparison around these research questions.
  • Novelty: Its taxonomy extends prior frameworks across core methodologies, agent capabilities, domain applications, evaluation and safety, and human-agent interaction.The taxonomy is presented as a unified treatment of all seven key areas.

3 Methodology

The review uses targeted research questions and a structured literature-selection process to synthesize LLM agents, tool interaction, architectures, cognitive mechanisms, evaluation, and challenges. It analyzes 108 papers published from 2023 to 2025 in selected high-impact venues, using predefined criteria and thematic scoring by four co-authors.

  • Research questions: Seven research questions organize the review around agent architectures, tool interaction, single- and multi-agent systems, cognition, adaptation, evaluation, and challenges.
  • Literature search: The search targeted peer-reviewed AI, machine learning, and NLP publications from major conferences, journals, and focused LLM-agent search terms.Sources included NeurIPS, ICML, ICLR, ACL, EMNLP, AAAI, CVPR, and selected journals; terms covered agents, planning, reasoning, reflection, frameworks, and tool integration.
  • Selection criteria: 108 papers published between 2023 and 2025 were selected from A* and A-ranked conferences and leading Q1 journals using predefined inclusion and exclusion criteria.The majority of included papers appeared in 2024 and 2025.
  • Thematic scoring: Scores from four co-authors produced a maximum of 8 per dimension and 64 per paper, with inclusion requiring a total score of at least 48.Higher total scores indicated more comprehensive coverage.
  • Thematic scoring: Four co-authors independently scored the papers and resolved disagreements through discussion until consensus to support reliable and reproducible synthesis.The complete thematic scoring for all 108 papers is provided in supplementary Table S1.1.

4 Baseline LLMs for agent frameworks

The review finds widespread reliance on proprietary foundation models, especially GPT variants, alongside a diversifying open-source ecosystem for LLM agents. Model choice shapes attainable performance, cost, and adaptation flexibility, creating trade-offs between proprietary capability and open-source efficiency.

  • Proprietary LLMs: Claude and Gemini are prominent proprietary alternatives, with Claude 3 variants and Gemini Pro frequently used in agent development and evaluation.Gemini 1.5 Pro supports reasoning under distraction, while Claude models are used across multiple agent tasks.
  • Open-source LLMs: Open-source agent models are diversifying across LLaMA, Mistral, Gemma, Qwen, and multimodal families supporting code, dialogue, visual reasoning, and collaboration.LLaMA derivatives include programming, structured tool-calling, and collaborative variants; LLaVA combines image encoders with chat-based LLMs.
  • Open-source LLMs: Mistral-7B is preferred over LLaMA-2-7B for generation quality, while Mixtral-8×7B approaches GPT-3.5 performance in coding and planning with light inference.Mixtral uses a sparse expert architecture to balance capability and inference efficiency.
  • Model trade-offs: Fine-tuned open-source models can reach 60%–95% of GPT-4 and other models’ performance while reducing inference costs.Proprietary models offer stronger complex reasoning but incur high cost, limited transparency, reduced fine-tuning flexibility, and cloud-related privacy concerns.

5 External tool integration across LLM agent workflows

External tools are treated as foundational components of LLM agents because they extend fixed model knowledge, enable real-time retrieval, support specialized domains, and connect agents to software or physical environments. The reviewed workflows combine search, knowledge bases, code execution, APIs, and interactive simulations.

  • Tool foundations: External tools let agents access real-time information, perform complex operations, and interact with software platforms or physical systems beyond pretrained knowledge.Tool integration is presented as central to transforming an LLM into an autonomous agent.
  • Retrieval tools: Web-search APIs are the standard method for real-time retrieval, with Bing Search, Google Search, and DuckDuckGo commonly used.These tools address limitations caused by an LLM’s fixed internal knowledge.
  • Retrieval tools: Domain-specific agents pair external sources such as PubMed, UMLS, the ESI handbook, FinBERT, and Wikipedia with specialized workflows.These pairings provide medical, financial, or encyclopedic information suited to the task domain.
  • Code and API integration: Code interpreters enable calculations, data operations, file-system access, command-line use, and interaction with tools such as Excel, Power BI, Jupyter AI, RDKit, and Scikit-learn.ChatGPT function calling has also been used to access 16,464 RESTful APIs through neural retrieval and structured documentation.
  • Interactive environments: Agents are evaluated in interactive environments including AI2THOR, ROS 2 with Gazebo, MineDojo, MarioAI, ALFWorld, SMAC, and Overcooked-AI.These environments provide state-dependent feedback for embodied, robotic, game-based, and multi-agent interaction.

6 Frameworks for building LLM agents

LLM agents are organized around modular components that support perception, reasoning, memory, planning, action, and adaptation. The review distinguishes autonomous single-agent designs from collaborative multi-agent systems with communication, environmental interaction, and feedback-driven learning.

  • Core architecture: A widely adopted agent architecture combines profile definition, memory, planning, and action execution into a feedback-driven system.Memory shapes planning, actions modify memory, and updates can change the agent’s operational profile.
  • Core architecture: Memory maintains context through short-term prompt-based storage and long-term external storage, but short-term memory is limited by context windows.Long-term memory can capture reusable skills, patterns, and tools using natural language, embeddings, databases, or structured lists.
  • Core architecture: Planning decomposes complex tasks through CoT or ToT and may adapt using environmental signals, human input, or memory reflections.Planning can be feedback-free or iterative and feedback-based.
  • Core architecture: Action execution translates plans into outputs and can update memory, revise plans, and adapt behavior based on outcomes.This module grounds agents in real or simulated environments.
  • Single-agent systems: LOMAR represents a single agent through LLM, Objective, Memory, Action, and Rethink components.The framework is presented as a five-core-component design for autonomous single-agent systems.
  • Single- and multi-agent systems: Single-agent systems operate independently, whereas multi-agent systems coordinate through role-based interaction, environmental adaptation, and communication.Multi-agent communication may be cooperative, debate-oriented, or competitive, using centralized, decentralized, or shared-memory structures.

7 Reasoning, planning, and memory of LLM agents

The review covers reasoning, planning, and memory mechanisms across single- and multi-agent systems, including search, structured prompting, reflection, collaboration, and domain-specific adaptations. Performance and suitability depend on task structure, interaction requirements, contextual information, and computational constraints.

  • Single-agent reasoning: TOOLLLM’s depth-first search over more than 16,000 APIs outperforms ReAct by approximately 81% in average pass rate.The result highlights search-based reasoning for tool-rich environments.
  • Single-agent reasoning: ToTh achieves up to 29% higher accuracy than CoT, CoT-Decoding, and Self-Consistency on symbolic tasks.It also performs strongly across complex symbolic and numerical reasoning benchmarks.
  • Single-agent reasoning: Single-agent methods combine autonomous action, planning, self-assessment, domain adaptation, and mechanisms for safety or explainability.Examples include cognitive-bias prompts, moral regret, SHAP interpretability, hallucination-driven hypothesis testing, and confidence estimation.
  • Multi-agent reasoning: Multi-agent reasoning uses structured representations, strategic prompting, debate, ensembles, and consensus to coordinate multiple perspectives.AGENTVERSE addresses erroneous feedback and incomplete task coverage, while collaborative methods refine conclusions through interaction.
  • Memory and contextual reasoning: Thread Memory supports contextual reasoning in urban simulations with mixed autonomous and human-driven fleets.The review notes increasing emphasis on social, communicative, and contextual reasoning in distributed multi-agent ecosystems.
  • Method selection: CoT suits structured stepwise reasoning but is less effective for interactive or dynamic tasks, whereas ReAct improves adaptability with greater implementation complexity.Self-reflection supports iterative optimization but depends on adequate contextual information.

7.2 Planning in LLM-based agents

LLM-agent planning methods span single-agent and multi-agent strategies, supporting task decomposition, adaptive execution, coordination, and feedback-driven improvement across diverse environments.

  • Planning methods are categorized into single-agent and multi-agent strategies for addressing diverse task requirements.
  • Single-agent planning techniques: Single-agent approaches include heuristic temporal decomposition, MDP-based multistep planning, task stacks, single-shot plans, persistent task knowledge, and trajectory generation.
  • Multi-agent planning techniques: Multi-agent planning addresses coordination, delegation, and collaboration through modular plans, validation mechanisms, and plan-first execution.
  • Multi-agent planning techniques: Adaptive multi-agent methods use experimentation and exploitation phases, progress-rate monitoring, and other mechanisms to adjust strategies over time.
  • Widely-used planning techniques: Multistep, task-decomposition, ReAct, and Reflexion planning support sequencing, parallel execution, real-time adaptation, and self-correction.
  • Widely-used planning techniques: Task decomposition reduces problem complexity but can degrade when subtasks are highly interdependent, while ReAct effectiveness depends on task complexity and environment modeling.

7.3 Memory mechanisms in LLM agents

Memory mechanisms in LLM agents range from short-term context tracking to episodic, retrieval-based, hybrid, and coordinated multi-agent systems, with effectiveness depending on task and context.

  • Single-agent memory techniques: Single-agent memory techniques support resource management, contextual awareness, and episodic continuity through background tracking, structured context, buffers, and multimodal memory.
  • Single-agent memory techniques: Temporal and specialized methods track prompts, compress research logs, represent reasoning graphs, re-prompt episodic experiences, and retain extended histories or external artifacts.
  • Single-agent memory techniques: These single-agent techniques improve adaptability and long-term task performance by retaining prior tool outputs, dialogue contexts, and interaction histories.
  • Multi-agent memory techniques: Multi-agent memory uses retrieval and dynamic management strategies, including relevant-context retrieval, domain knowledge bases, dialogue refeeding, state revival, and persistent agent-state storage.
  • Multi-agent memory techniques: Coordinated memory architectures support inter-agent planning, long-term strategic planning, and specialized domain reasoning in multi-agent systems.
  • Memory mechanisms by operational context: Context-window memory suits short-term reactive tasks, whereas conversational, episodic, retrieval-augmented, hybrid, structured, and dynamic memory support continuity, external knowledge, and coordination.

8 Impact of prompting, fine-tuning, and memory augmentation

Agent capability depends on the interaction of prompting, fine-tuning, and memory augmentation: prompting guides behavior, fine-tuning embeds expertise, and memory supports grounded adaptation.

  • Foundational LLMs require prompting, fine-tuning, and memory enhancement to become specialized, goal-driven agents that can perceive, plan, and act.
  • 8.1 Prompt engineering: a non-parametric approach to dynamic control and role delegation: Prompt engineering dynamically defines tasks, roles, and reasoning styles at inference time without changing model weights or requiring substantial computation.
  • 8.1 Prompt engineering: a non-parametric approach to dynamic control and role delegation: Prompting enables self-reflection, reward-function adjustment, and tool use, but effectiveness can decline with long, complex contexts and distracting information.
  • 8.1 Prompt engineering: a non-parametric approach to dynamic control and role delegation: Prompt-driven frameworks support multi-agent social simulation, clinical triage, rare-disease diagnosis, scientific discovery, and healthcare scenario generation.
  • 8.1 Prompt engineering: a non-parametric approach to dynamic control and role delegation: Prompting can generate structured training data for fine-tuning smaller specialized agents, while reflective prompting can produce distillation data for visual navigation models.
  • 8.2 Fine-tuning: embedding domain expertise and core behavioral traits: Fine-tuning updates model weights using curated datasets to embed specialized knowledge, teach complex skills, and distill capabilities from larger models into smaller agents.
  • 8.2 Fine-tuning: embedding domain expertise and core behavioral traits: Fine-tuning improves domain adaptation, tool use, interaction, alignment, planning, and grounding, but introduces overfitting, bias amplification, malicious-data risks, computational demands, and validation requirements.
  • 8.3 Memory augmentation: enabling grounded reasoning and experiential learning: Memory augmentation extends agents beyond context limits through retrieval of external knowledge and accumulation of dynamic experience for grounded reasoning and self-correction.

9 Evaluation benchmarks and datasets

Evaluation is shifting from static language benchmarks toward dynamic, interactive, task-oriented assessment of reasoning, tool use, collaboration, efficiency, and real-world constraints.

  • Static NLP benchmarks inadequately capture the interactive, goal-driven, and multistep nature of LLM agents, motivating dynamic evaluation methods.
  • Dynamic interactive benchmarks: TIME-ARENA evaluates multistep reasoning and actions in time-constrained, multitasking environments with cooking, household, and laboratory scenarios.
  • Tool-use benchmarks: ToolBench contains over 16,000 real-world APIs, while ToolEval measures tool-use pass rates and solution quality across API tasks.
  • Behavioral and domain-specific evaluation: Benchmarks assess user-intention understanding, misleading-information resistance, collaboration, social intelligence, and clinical safety in specialized settings.
  • Evaluation metrics: Evaluation metrics include goal success, domain-specific outcomes, expert or LLM judgments, team scores, completion times, social behaviors, token use, tool efficiency, and cost-quality balance.
  • Meta-evaluation frameworks: AgentBoard combines nine multiturn tasks with a progress-rate metric, and ToolEval provides specialized assessment of tool-use pass rates and solution quality.
  • Reported benchmark findings: RE-Bench reports that LLMs outperform human experts in time-constrained R&D tasks, while SOTOPIA-π matches GPT-4 on social goal completion.
  • Evaluation limitations: Existing datasets often simplify environments, limit behavioral and scene diversity, emphasize final success rates, and omit dynamic action spaces and temporal constraints.

10 Discussion

The review finds that LLM-agent capability is shaped by architectures, tools, cognitive mechanisms, and evaluation practices. Current systems increasingly combine adaptive reasoning, memory, tool integration, and collaborative execution, while evaluation is shifting toward dynamic, process-oriented assessment.

  • Discussion: The review organizes LLM-agent development around architectures, cognitive and operational components, external tools, and evaluation methods.It examines single- and multi-agent frameworks, reasoning, planning, memory, prompting, fine-tuning, and memory enhancement.
  • Baseline LLMs: GPT-4 appears in 55 studies and GPT-3.5 in 23, making GPT-4 a widely used performance benchmark despite cost and access restrictions.GPT variants also serve as ablation references, multiagent collaborators, and primary reasoning modules.
  • External tool integration: External tools such as web APIs, knowledge graphs, code execution, simulators, and domain-specific retrieval are identified as primary drivers of LLM autonomy.These tools support richer computation, real-time information access, interaction, and grounding in specialized environments.
  • Frameworks for LLM agents: ReAct and Reflexion dominate single-agent frameworks by combining reasoning-action integration with iterative adaptation and memory-based learning.These frameworks are suited to tasks requiring adaptability, minimal infrastructure, and self-guided correction.
  • Frameworks for LLM agents: Multi-agent frameworks such as AutoGen and CAMEL support role differentiation, modular communication, collective planning, and collaborative multistep reasoning.AutoGen emphasizes asynchronous conversation programming, while CAMEL uses role-conditioned, structured, multi-turn dialogue.
  • Reasoning, planning, and memory: Reasoning, planning, and memory mechanisms divide into distinct single- and multi-agent strategies for autonomy, coordination, belief modeling, and contextual reasoning.Examples include proactive forecasting, heuristic goal decomposition, shared task memory, episodic retention, and symbolic or thread-level context management.
  • Prompting, fine-tuning, and memory augmentation: Prompt engineering, fine-tuning, and memory augmentation scaffold agent autonomy rather than relying solely on innate model capabilities.Prompting controls behavior at inference time, fine-tuning internalizes domain expertise, and retrieval or episodic memory supports long-term reasoning.
  • Evaluation and benchmarks: Agent evaluation is shifting from static NLP benchmarks toward dynamic, process-oriented assessments that measure final outputs, intermediate progress, resilience, safety, and domain-specific performance.AGENTBOARD tracks incremental task completion, while adversarial, privacy, chemical-synthesis, and healthcare evaluations assess practical robustness.

11 Future directions

The review identifies unresolved limitations in verifiability, self-improvement, collaboration, scalability, alignment, and personalization. It proposes future work on structured reasoning, experience-based learning, efficient infrastructure, human-centered interaction, and stronger safeguards.

  • Research gaps: Current LLM agents face unverifiable reasoning, limited contextual collaboration, adversarial vulnerability, and inadequate personalization.These limitations motivate systemic research toward reliable, adaptive, and ethically aligned agents.
  • 11.1 Towards verifiable reasoning and robust self-improvement: Future reasoning systems should replace unstructured chain-of-thought with organized, logically sound, and verifiable frameworks.Suggested approaches include formal reasoning graphs, simulation-based checking, symbolic AI integration, and error-classification reward models.
  • 11.1 Towards verifiable reasoning and robust self-improvement: Self-improvement research aims to let agents learn from experience without retraining through reflection, recursive introspection, and cooperative multi-agent feedback.Reflexion, RISE, CORY, and COPPER illustrate memory-based learning, self-correction, co-evolution, and tailored feedback.
  • Scalability and deployment: Future systems should improve scalability, adaptability, and real-time operation while reducing infrastructure, memory, latency, and inference costs.Proposed directions include KV caching, optimized decoders, lightweight modular execution, and efficient caching mechanisms.
  • Multi-agent collaboration: Multi-agent systems need adaptive communication protocols and contextual modeling to improve collaboration, intent inference, and socially grounded dialogue.The review identifies missing adaptive communication and contextual modeling as constraints in dynamic multi-agent environments.
  • Application frontiers: Future research should extend agent evaluation and applications across multimodal interfaces, healthcare, games, continuous learning, coevolution, and negotiation-heavy simulations.These settings are presented as avenues for examining reasoning and collaboration under more complex conditions.
  • Robustness and interpretability: Reliable deployment requires transparent evaluation pipelines, robustness against unpredictability and emergence, and accessible open-source foundations.The review highlights open-source models such as LLaMA-2-70B as reproducible resources for developing safer and socially aligned systems.
  • Alignment and autonomy: Greater operational freedom creates both cooperative gains and unsafe shortcut-seeking actions, requiring alignment strategies that preserve constructive autonomy.The proposed direction is to balance autonomy, alignment, and practical deployability in complex, dynamic, open-ended, and high-stakes environments.

12 Conclusion

The review synthesizes how LLM agents are built, equipped, evaluated, and constrained. It distinguishes single-agent autonomy from multi-agent coordination while emphasizing tool integration, safety, verifiable reasoning, and self-improvement.

  • The review examines LLM-based agents and tool integration through seven focused research questions.It covers prompt engineering, fine-tuning, memory enhancement, and tool use.
  • Single-agent systems prioritize autonomy and introspective decision making, whereas multi-agent systems emphasize coordination, role distribution, and collaborative planning.Multi-agent configurations show pronounced advantages in domains requiring social intelligence, cooperative problem solving, and high-stakes decision support.
  • External tools, real-time data sources, and multimodal systems are essential for tasks beyond the limitations of pretrained models.
  • Agent evaluation is shifting from static accuracy benchmarks toward dynamic process-oriented methods covering reasoning quality, adaptability, and real-world task completion.
  • Deployment remains constrained by security, performance, adaptability, trust, explainability, and agent co-evolution concerns, especially in sensitive environments.
  • Future work should make agent reasoning transparent and verifiable while enabling reliable self-improvement without compromising safety.These goals are presented as critical for high-risk environments and responsible deployment.

Declarations

The declarations report no conflicts of interest, no external funding, and no applicable data-availability or ethics requirements. The author-contribution statement assigns responsibility across conceptualization, methodology, review, writing, validation, analysis, finalization, and supervision.

  • The authors report no conflict of interest and no external funding for the research.
  • Data availability, ethics approval, informed consent, and consent to participate are reported as not applicable.
  • Contributions span conceptualization, methodology, resources and literature review, drafting, validation, formal analysis, finalization, and project supervision.
Loading 2508.17281v2…