从语言到行动：大语言模型作为自主智能体与工具使用者的综述

Sadia Sultana Chowa1, Riasad Alvi2, Subhey Sadi Rahman2, Md Abdur Rahman2, Mohaimenul AzamKhan Raiaan2, Md Rafiqul Islam3, Mukhtar Hussain3, Sami Azam1Department of Computer Science and Engineering, Daffodil International University, Dhaka-1341, Bangladesh2Department of Computer Science and Engineering, United International University, Dhaka 1212, Bangladesh3Faculty of Science and Technology, Charles Darwin University, Casuarina, NT 0909, Australia*Corresponding Author: sami.azam@cdu.edu.auAbstractprompting [3], chain-of-thought (CoT) prompting [4], and self- The pursuit of human-level artificial intelligence (AI) has sig-nificantly advanced the development of autonomous agentsand Large Language Models (LLMs). LLMs are now widelyutilized as decision-making agents for their ability to interpretinstructions, manage sequential tasks, and adapt through feed-back. This review examines recent developments in employingLLMs as autonomous agents and tool users and comprisesseven research questions. We only used the papers publishedbetween 2023 and 2025 in conferences of the A* and A rankand Q1 journals. A structured analysis of the LLM agents’architectural design principles, dividing their applications intosingle-agent and multi-agent systems, and strategies for inte-grating external tools is presented. In addition, the cognitivemechanisms of LLM, including reasoning, planning, and mem-ask prompting [5] demonstrated how the potential of LLMscould be improved through smart prompting and input pat-tern design. Beyond conventional natural language processing(NLP) tasks, LLMs are now serving as autonomous agents andintelligent tools. They are embedded into increasingly complexworkflows where they perform planning, decision making, andtool interaction in various real-world applications, includingresearch assistance [6], software development [7], drug discov-eries [8], multi-robot systems [9], clinical support [10], gamesimulation [11] and scientific simulations [12].LLMs as agents can observe their environment, make deci-sions, and take actions. Within this paradigm, single-agentLLM systems have demonstrated promising performance indecision-making tasks. Single-agent systems such as Reflex- ory, and the impact of prompting methods and fine-tuningprocedures on agent performance are also investigated. Fur-thermore, we evaluated current benchmarks and assessmentprotocols and have provided an analysis of 68 publicly avail-able datasets to assess the performance of LLM-based agentscan operate in decision loops that involve planning, mem-ory, and tool use. However, they often struggle in dynamicenvironments that require simultaneous context tracking, ex-ternal memory integration, and adaptive tool usage [16, 17].To address these limitations, the concept of multi-agent LLM agents. Finally, we have discussed ten future research direc-tions to overcome these gaps.single agent can manage. Through structured communication,reflective reasoning, and explicit role assignments in simulatedarXiv:2508.17281v1 [cs.CL] 24 Aug 2025 Keywords:Large Language Models;Multi-Agents;Reasoning; Evaluation; Generative AI 1Introduction Large language models (LLMs) have become central in artifi-cial intelligence (AI) research due to their strong human-likeability to understand, generate, and reason in natural lan-guage [1, 2].LLMs were used primarily as tools to serveas text generators or understanding modules within a largerapplication.However, further techniques such as few-shotMoreover, LLMs as agents and tools now demonstrate mas-sive potential in AI, and the demand to understand theirevolving roles has intensified. Therefore, a systematic reviewof its recent advancement, a discussion of the remaining gaps,and a research direction for future advancements are essential to advance the field.With this focus, this survey providesa comprehensive and structured overview of current capabil-ities and system designs.We investigate the architecturalfoundations that enable agent-like behavior in LLMs, analyzehow they interact with external tools, discuss the key limita-tions of current approaches, and highlight the remaining open benchmarks for LLM agents and tool users. •We identify fundamental challenges, including alignment,reliability, and generalization, and outline promising re- The rest of the review is organized as follows.Section 2 presents related works, identifying gaps in existing surveysand situating our contribution. Section 3 outlines the method- Our key contributions are summarized as follows. ology, including research questions, selection criteria, andsearch strategies. Section 4 explores the baseline LLMs usedin agentic LLM systems.Section 5 focuses on tool integra-tion in LLM workflows. Section 6 reviews the frameworks forconstructing single-agent and multi-agent systems. Section 7investigates the reasoning, planning, and memory capabilities •We conduct a comprehensive review of recent advance-ments in using LLMs as agents and