Modern Multi-Agent Systems: Advances, Architectures, and Challenges 【Synthesis of Recent Research 】-CSDN博客

本文链接：https://blog.csdn.net/universsky2015/article/details/148645652

Modern Multi-Agent Systems: Advances, Architectures, and Challenges

Synthesis of Recent Research (2023-2025)

Executive Summary

Recent research in Multi-Agent Systems (MAS) reveals rapid evolution driven by large language models (LLMs), with breakthroughs in collaborative frameworks, security protocols, and specialized applications (e.g., AI art, blockchain). Key trends include:

Terminological debates questioning “Agentic AI” as a rebranding of classical MAS concepts (Botti, 2025).
Novel architectures like the Athenian Academy’s seven-layer model for complex task decomposition (Zhai et al., 2025).
Critical vulnerabilities in MAS (e.g., backdoor attacks) and defenses like mutual reasoning (Fan & Li, 2025).
Persistent challenges in task allocation, context management, and system stability (Han et al., 2024).

1. Foundations and Terminological Tensions

1.1 Core MAS Principles

MAS enable autonomous agents to interact, cooperate, or compete in shared environments. Foundational properties include:

Autonomy, reactivity, proactivity, and social capability (Wooldridge & Jennings).
Classical architectures: Reactive agents → Belief-Desire-Intention (BDI) models (Botti, 2025).

1.2 The “Agentic AI” Debate

Botti (2025) argues: Terms like “Agentic AI” and “Multiagentic” misrepresent established MAS research.
- Critique: LLM-based agents are evolutional, not revolutionary; MAS already addressed communication, negotiation, and trust.
- Risk: Ignoring decades of work (e.g., FIPA standards) leads to redundant solutions (Botti, 2025).

Recommendation: Adopt classical MAS terminology and integrate existing knowledge (e.g., agreement technologies) into LLM-based agents.

2. Architectural Innovations

2.1 The Athenian Academy Framework (Zhai et al., 2025)

A seven-layer architecture for MAS in AI art creation:

Multi-agent collaboration
Single-agent multi-role playing
Single-agent multi-scene traversal
Single-agent multi-capability incarnation
Agents sharing LLMs for identical goals
Single-agent using multiple LLMs for identical goals
Multi-agent synthesis for unified outputs
Applications: Enhanced cross-scene adaptation and model fusion in art generation.

2.2 AutoGenesisAgent (2024)

Self-generating MAS for complex tasks via LLM orchestration.
Key feature: Dynamic agent creation/retirement based on task requirements.

2.3 Hybrid and Hierarchical MAS

Hybrid MAS: Integrates continuous/discrete-time agents for consensus (2015).
Hierarchical MAS: Time-varying active agent sets with stability guarantees (2022).

3. Security and Trust Challenges

3.1 Backdoor Attacks in MAS (Fan & Li, 2025)

Threat: Malicious triggers compromise agents during inference (e.g., poisoned financial advice).
Defense: PeerGuard — agents mutually audit reasoning chains:
- Step 1: Require explicit reasoning traces.
- Step 2: Flag logic-output inconsistencies.
- Effectiveness: 92% detection accuracy in GPT/Llama3-based MAS with low false positives.

3.2 Trust and Coordination

Relational networks impact learning efficiency (value-based factorization; 2023).
Ontology-based feedback improves shop-floor control (manufacturing MAS; 2023).

Actionable Insight: Deploy mutual reasoning defenses in high-stakes MAS (e.g., healthcare, finance).

4. Emerging Challenges and Open Problems

4.1 Core Unsolved Issues (Han et al., 2024)

Task allocation: Optimizing agent specialization (e.g., debate vs. reflection roles).
Context management: Aligning layered contexts (global, agent-specific, shared knowledge).
Memory systems: Lack of frameworks for cross-agent memory access/learning.
Reasoning robustness: Ensuring debates refine (not derail) outputs.

4.2 Stability and Scalability

Evolving MAS: Stability risks in highly connected, nonlinear systems (2011).
Molecular communication MAS: Scalability limits in nanorobot networks (2023).
Multi-agent reinforcement learning (MARL): Collaborative emergence ≠ improved performance (2018).

5. Applications and Future Directions

5.1 High-Impact Domains

AI Art: Athenian Academy enables complex co-creation (Zhai et al., 2025).
Blockchain:
- Agents as tools: Audit smart contracts.
- Agents as user representatives: Autonomous node operations (Han et al., 2024).
Machine Translation: Single/multi-agent systems for multilingual communication (2025).
Robotics: Reciprocal collision avoidance via MARL (2019).

5.2 Research Priorities

Standardize terminology to bridge classical MAS and LLM-based agents.
Develop hybrid security: Combine mutual reasoning with encryption (e.g., Agent TCP/IP; 2025).
Address MAS stability: Formal methods for consensus in hybrid/dynamic systems.
Optimize LLM multi-agents: Reduce hallucinations via BDI-inspired constraints.

References

Botti, V. (2025). Agentic AI and Multiagentic: Are We Reinventing the Wheel? arXiv:2506.01463. PDF
Fan, F., & Li, X. (2025). PeerGuard: Defending MAS Against Backdoor Attacks. arXiv:2505.11642v2. PDF
Han, S. et al. (2024). LLM Multi-Agent Systems: Challenges and Open Problems. arXiv:2402.03578v2. PDF
Zhai, L. et al. (2025). The Athenian Academy: A MAS Architecture. arXiv:2504.12735v2. PDF
VAIN Framework (2017). Attentional Multi-agent Predictive Modeling. arXiv:1706.06122v2. PDF

Report synthesizes 17 ArXiv papers (2017–2025), prioritizing recent advances (>2023) and thematic cohesion. Citations include first-author names + year for traceability.
ENDOFTEXTSTREAM

Gathered information: Title: VAIN: Attentional Multi-agent Predictive Modeling
Date: 2017-06-19
PDF Url: http://arxiv.org/pdf/1706.06122v2

Multi-agent predictive modeling is an essential step for understanding
physical, social and team-play systems. Recently, Interaction Networks (INs)
were proposed for the task of modeling multi-agent physical systems, INs scale
[…]

Title: The Athenian Academy: A Seven-Layer Architecture Model for Multi-Agent Systems
Date: 2025-04-17
PDF Url: http://arxiv.org/pdf/2504.12735v2

This paper proposes the “Academy of Athens” multi-agent seven-layer
framework, aimed at systematically addressing challenges in multi-agent systems
(MAS) within artificial intelligence (AI) art creation, such as collaboration
[…]

Title: Agentic AI and Multiagentic: Are We Reinventing the Wheel?
Date: 2025-06-02
PDF Url: http://arxiv.org/pdf/2506.01463v1

The terms Agentic AI and Multiagentic AI have recently gained popularity in
discussions on generative artificial intelligence, often used to describe
autonomous software agents and systems composed of such agents. However, the
[…]

Title: Are AI agents the new machine translation frontier? Challenges and opportunities of single- and multi-agent systems for multilingual digital communication
Date: 2025-04-17
PDF Url: http://arxiv.org/pdf/2504.12891v1

The rapid evolution of artificial intelligence (AI) has introduced AI agents
as a disruptive paradigm across various industries, yet their application in
machine translation (MT) remains underexplored. This paper describes and
[…]

Title: Agent TCP/IP: An Agent-to-Agent Transaction System
Date: 2025-01-08
PDF Url: http://arxiv.org/pdf/2501.06243v1

Autonomous agents represent an inevitable evolution of the internet. Current
agent frameworks do not embed a standard protocol for agent-to-agent
interaction, leaving existing agents isolated from their peers. As intellectual
[…]

Title: AutoGenesisAgent: Self-Generating Multi-Agent Systems for Complex Tasks
Date: 2024-04-25
PDF Url: http://arxiv.org/pdf/2404.17017v1

The proliferation of large language models (LLMs) and their integration into
multi-agent systems has paved the way for sophisticated automation in various
domains. This paper introduces AutoGenesisAgent, a multi-agent system that
[…]

Title: LLM Multi-Agent Systems: Challenges and Open Problems
Date: 2024-02-05
PDF Url: http://arxiv.org/pdf/2402.03578v2

Title: PeerGuard: Defending Multi-Agent Systems Against Backdoor Attacks Through Mutual Reasoning
Date: 2025-05-16
PDF Url: http://arxiv.org/pdf/2505.11642v2

Multi-agent systems leverage advanced AI models as autonomous agents that
interact, cooperate, or compete to complete complex tasks across applications
such as robotics and traffic management. Despite their growing importance,
[…]

Title: Impact of Relational Networks in Multi-Agent Learning: A Value-Based Factorization View
Date: 2023-10-19
PDF Url: http://arxiv.org/pdf/2310.12912v1

Effective coordination and cooperation among agents are crucial for
accomplishing individual or shared objectives in multi-agent systems. In many
real-world multi-agent systems, agents possess varying abilities and
[…]

Title: Consensus of hierarchical multi-agent systems with a time-varying set of active agents
Date: 2022-12-01
PDF Url: http://arxiv.org/pdf/2212.00455v1

Time-varying hierarchical multi-agent systems are common in many
applications. A well-known solution to control these systems is to use state
feedback controllers that depend on the adjacency matrix to reach consensus.
[…]

Title: Stability of Evolving Multi-Agent Systems
Date: 2011-11-30
PDF Url: http://arxiv.org/pdf/1111.7033v1

A Multi-Agent System is a distributed system where the agents or nodes
perform complex functions that cannot be written down in analytic form.
Multi-Agent Systems are highly connected, and the information they contain is
[…]

Title: Multi-Agent Training for Pommerman: Curriculum Learning and Population-based Self-Play Approach
Date: 2024-06-30
PDF Url: http://arxiv.org/pdf/2407.00662v2

Pommerman is a multi-agent environment that has received considerable
attention from researchers in recent years. This environment is an ideal
benchmark for multi-agent training, providing a battleground for two teams with
[…]

Title: Ontology-Based Feedback to Improve Runtime Control for Multi-Agent Manufacturing Systems
Date: 2023-09-18
PDF Url: http://arxiv.org/pdf/2309.10132v1

Improving the overall equipment effectiveness (OEE) of machines on the shop
floor is crucial to ensure the productivity and efficiency of manufacturing
systems. To achieve the goal of increased OEE, there is a need to develop
[…]

Title: Consensus of Hybrid Multi-agent Systems
Date: 2015-12-10
PDF Url: http://arxiv.org/pdf/1512.03189v1

In this paper, we consider the consensus problem of hybrid multi-agent
system. First, the hybrid multi-agent system is proposed which is composed of
continuous-time and discrete-time dynamic agents. Then, three kinds of
[…]

Title: Measuring collaborative emergent behavior in multi-agent reinforcement learning
Date: 2018-07-23
PDF Url: http://arxiv.org/pdf/1807.08663v1

Multi-agent reinforcement learning (RL) has important implications for the
future of human-agent teaming. We show that improved performance with
multi-agent RL is not a guarantee of the collaborative behavior thought to be
[…]

Title: Multi-agent Policy Optimization with Approximatively Synchronous Advantage Estimation
Date: 2020-12-07
PDF Url: http://arxiv.org/pdf/2012.03488v3

Cooperative multi-agent tasks require agents to deduce their own
contributions with shared global rewards, known as the challenge of credit
assignment. General methods for policy based multi-agent reinforcement learning
[…]

Title: Reciprocal Collision Avoidance for General Nonlinear Agents using Reinforcement Learning
Date: 2019-10-24
PDF Url: http://arxiv.org/pdf/1910.10887v2

Finding feasible and collision-free paths for multiple nonlinear agents is
challenging in the decentralized scenarios due to limited available information
of other agents and complex dynamics constraints. In this paper, we propose a
[…]

Title: Stability analysis for large-scale multi-agent molecular communication systems
Date: 2023-11-12
PDF Url: http://arxiv.org/pdf/2311.06730v2

Molecular communication (MC) is recently featured as a novel communication
tool to connect individual biological nanorobots. It is expected that a large
number of nanorobots can form large multi-agent MC systems through MC to
[…]

Title: Reaching a Consensus in Networks of High-Order Integral Agents under Switching Directed Topology
Date: 2013-04-15
PDF Url: http://arxiv.org/pdf/1304.3972v1

Consensus problem of high-order integral multi-agent systems under switching
directed topology is considered in this study. Depending on whether the agent’s
full state is available or not, two distributed protocols are proposed to
[…]

Title: Contract-based Design and Verification of Multi-Agent Systems with Quantitative Temporal Requirements
Date: 2024-12-17
PDF Url: http://arxiv.org/pdf/2412.13114v1

Quantitative requirements play an important role in the context of
multi-agent systems, where there is often a trade-off between the tasks of
individual agents and the constraints that the agents must jointly adhere to.
[…]

Agentic AI and Multiagentic: Are We Reinventing the Wheel?

Vicent Botti

Valencian Research Institute for Artificial Intelligence (VRAIN) Universitat Politècnica de València (UPV) vbotti@vrain.upv.es

Valencian Graduated School and Research Network of Artificial Intelligence (ValgrAI)

Abstract

The terms " Agentic AI " and " Multiagentic AI " have recently gained popularity in discussions on generative artificial intelligence, often used to describe autonomous software agents and systems composed of such agents. However, the use of these terms confuses these buzzwords with wellestablished concepts in AI literature: intelligent agents and multi-agent systems . This article offers a critical analysis of this conceptual misuse. We review the theoretical origins of “agentic” in the social sciences (Bandura, 1986) and philosophical notions of intentionality (Dennett, 1971), and then summarise foundational works on intelligent agents and multi-agent systems by Wooldridge, Jennings and others. We examine classic agent architectures -from simple reactive agents to Belief-Desire-Intention (BDI) models -and highlight key properties (autonomy, reactivity, proactivity, social capability) that define agency in AI. We then discuss recent developments in large language models (LLMs) and agent platforms based on LLMs, including the emergence of LLM-powered AI agents and open-source multi-agent orchestration frameworks. We argue that the term “AI Agentic” is often used as a buzzword for what are essentially AI agents, and “AI Multiagentic” for what are multi-agent systems. This confusion overlooks decades of research in the field of autonomous agents / multi-agent systems. The article advocates for scientific and technological rigour and the use of established terminology from the state of the art in AI, incorporating the wealth of existing knowledge -including standards for multi-agent system platforms, communication languages and coordination/cooperation algorithms, agreement technologies (automated negotiation, argumentation, virtual organisations, trust, reputation, etc.) -into the new and promising wave of LLM-based AI agents, so as not to end up reinventing the wheel .

Introduction

Recent advances in generative AI , particularly large language models (LLMs), have led to a resurgence of interest in autonomous software agents : AI systems that can autonomously perceive their environment, reason, and act to meet their design goals to perform tasks on behalf of users. Visionaries such as Bill Gates predict that, in the near future, we will all have a personal AI assistant “far superiror to current technology,” capable of responding to natural language requests and performing various tasks by understanding the user’s goals [6]. Gates notes that this software, which he and others have imagined for decades, is finally becoming practical thanks to advances in AI, and that such " agents " could revolutionize our interaction with computers. In parallel, some industrial discourse, not always; there are many cases where the terminology is respected, has introduced new terms such as " Agentic AI " to describe AI systems endowed with autonomy and proactive decision-making. Technical blogs and articles differentiate between “AI agents” and “Agentic AI,” presenting the latter as a general framework where multiple agents operate. Similarly, the term " Multiagentic " is sometimes used for systems with multiple interacting agents, analogous to classic multiagent systems .

This emergence of ‘agentic’ terminology in AI raises critical questions. Are these genuinely new concepts, or simply new labels for already established ideas in autonomous agents research? The field of intelligent agents and multi-agent systems (MAS) has a rich history dating back decades, with its own theoretical

foundations, architectures, and even dedicated conferences, journals and scientific associations of researchers in the field. Before embracing the hype of ‘agent ic AI,’ it is important to examine whether the term is being misapplied as a buzzword for capabilities long understood in AI -In other words, are we reinventing the wheel by equating agentic AI with intelligent agents and multiagentic systems with multiagent systems ?

The Athenian Academy: A Seven-Layer Architecture Model for Multi-Agent Systems

Lidong Zhai ∗ a& , Zhijie Qiu †∥ a , Lvyang Zhang ∗‡ , Jiaqi Li ∗‡ , Yi Wang , Wen Lu § ∗‡ Xizhong Guo ∗‡ , Ge Sun ¶

∗ Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China ‡ School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China † Tianjin Academy of Fine Arts, AI Art Institute, Tianjin, China ∥ Central Academy of Fine Arts, Institute of Science and Technology in Art, Beijing, China § Central Academy of Fine Arts, Beijing, China

¶ WaytoAGI, China

Abstract -This paper proposes the ‘Academy of Athens’ multi-agent seven-layer framework, aimed at systematically addressing challenges in multi-agent systems (MAS) within artificial intelligence (AI) art creation, such as collaboration efficiency, role allocation, environmental adaptation, and task parallelism. The framework divides MAS into seven layers: multi-agent collaboration, single-agent multi-role playing, single-agent multiscene traversal, single-agent multi-capability incarnation, different single agents using the same large model to achieve the same target agent, single-agent using different large models to achieve the same target agent, and multi-agent synthesis of the same target agent. Through experimental validation in art creation, the framework demonstrates its unique advantages in task collaboration, cross-scene adaptation, and model fusion. This paper further discusses current challenges such as collaboration mechanism optimization, model stability, and system security, proposing future exploration through technologies like metalearning and federated learning. The framework provides a structured methodology for multi-agent collaboration in AI art creation and promotes innovative applications in the art field.

keywords: Athenian Academy; Multi-Agent Systems; Artistic Creation;

I. INTRODUCTION

In the grand scene of Raphael’s painting ‘The School of Athens’, we see philosophers and thinkers engaged in passionate discussion, each striving to push the boundaries of human understanding and creativity. The painting is set on the walls of the sacred Academy of Athens, serving as a symbol of the intersection between wisdom and artistic expression, showcasing the collision, evolution, and generational transmission of ideas. Within this temple of knowledge, figures like Plato, Aristotle, and Socrates exchange profound insights, shaping the future of philosophical thought and artistic innovation. Just as these figures advanced the development of ideas through rigorous debate and creative exploration, modern artificial intelligence (AI) systems, particularly within the framework of multi-agent systems (MAS), are replicating this collaborative creativity through advanced computational methods.

This paper explores the intersection of multi-agent systems (MAS) and AI-driven art creation, proposing a new approach to explore how agents can aggregate their powers to create

a These authors contributed equally to this work.

& Corresponding author: zhailidong@iie.ac.cn

Fig. 1: Raphael’s The School of Athens

art that transcends individual limitations. MAS is a complex system composed of multiple autonomous agents interacting within a shared environment, which has long been a research topic in the field of artificial intelligence. MAS enables distributed decision-making, parallel processing, and strong environmental adaptability, all of which are crucial for solving complex dynamic tasks.

In recent years, AI-based art creation has gradually gained prominence, particularly due to the powerful capabilities of generative models and machine learning. These models can generate, manipulate, and evaluate art in unprecedented ways, pushing the boundaries of creativity. However, just like the ancient philosophers in ‘The School of Athens’, AI agents must also learn to interact, collaborate, and debate within a shared intellectual space to create cohesive and innovative artistic works. This paper proposes a framework inspired by ‘The School of Athens’, called the ‘Academy of Athens Framework’, for multi-agent art creation. The framework introduces a seven-layer MAS collaboration approach aimed at addressing the challenges and potentials in multi-agent art creation.

The ‘Academy of Athens Framework’ divides the collaboration process into seven layers, starting from the perception and cognition of individual agents, gradually advancing to agent collaboration, role allocation, and multi-agent coordination, and ultimately achieving the holistic creation of art. This method, as a structured abstract perspective, is employed to understand and solve complex system problems. This multilevel design not only facilitates a deeper analysis of the various components of the MAS but also provides a flexible structure for AI agents to engage in detailed and dynamic artistic creation.

PeerGuard: Defending Multi-Agent Systems Against Backdoor Attacks Through Mutual Reasoning

1 st Falong Fan The Chinese University of Hong Kong, Shenzhen falongfan@link.cuhk.edu.cn

Abstract -Multi-agent systems leverage advanced AI models as autonomous agents that interact, cooperate, or compete to complete complex tasks across applications such as robotics and traffic management. Despite their growing importance, safety in multi-agent systems remains largely underexplored, with most research focusing on single AI models rather than interacting agents. This work investigates backdoor vulnerabilities in multiagent systems and proposes a defense mechanism based on agent interactions. By leveraging reasoning abilities, each agent evaluates responses from others to detect illogical reasoning processes, which indicate poisoned agents. Experiments on LLM-based multi-agent systems, including ChatGPT series and Llama 3, demonstrate the effectiveness of the proposed method, achieving high accuracy in identifying poisoned agents while minimizing false positives on clean agents. We believe this work provides insights into multi-agent system safety and contributes to the development of robust, trustworthy AI interactions. Our code is available in the link 1 at the footnote.

Index Terms -Multi-agent systems, Backdoor Defense, Large Language Models, Chain of Thoughts

I. INTRODUCTION

Multi-agent systems (MAS) use large language models (LLMs) as autonomous agents that interact to accomplish complex tasks across various applications [1, 2]. While their use is expanding, safety in multi-agent settings remains underexplored, with most research focusing on individual models rather than agent interactions. These systems inherit vulnerabilities from LLMs: pre-training on large-scale Internet data introduces harmful content such as bias and racism [3]. Besides, advanced features like in-context learning make attacks easier to execute. For example, poisoning attacks can occur at inference time via malicious prompts, bypassing the need to alter training data [4, 5]. Such vulnerabilities may propagate and intensify through agent interactions [6], making trustworthiness a growing concern [7].

Among the threats to multi-agent systems, we focus on backdoor attacks - an established and potent class of attacks in the AI community. These attacks exploit a predefined trigger to induce malicious behavior in one or more agents while preserving normal performance on clean inputs. The attack can propagate through agent interactions and influence the collective decision-making process. The widespread use of third-party LLM services, including APIs and prompt

2 nd Xi Li

University of Alabama at Birmingham

XiLiUAB@uab.edu engineering tools, further increases the attack surface: unregulated providers may embed malicious instructions in prompts without altering the model itself [8, 9]. For example, in a multi-agent financial assistant system, a poisoned agent could be triggered to recommend risky investments, misleading the other agent in the debate and ultimately influencing the final consensus toward harmful outcomes.

However, existing backdoor defense research largely focuses on single LLMs and addresses a limited set of attack types, with minimal exploration in MAS. For instance, [10] proposes detecting out-of-distribution words in the input to defend against textual backdoor attacks, but this approach is ineffective against attacks that do not rely on irregular trigger tokens. Similarly, [11] filters suspicious content from training data, which is impractical for most modern LLMs accessed only via APIs without training data visibility. [12] introduces a coordinator agent in MAS to detect jailbreak attacks, but focuses solely on malicious prompts and overlooks the deeper threat of compromised models. Other works [13, 14] study prompt injection propagation across agents without altering the underlying model. In contrast, our work investigates modellevel backdoors that embed malicious behaviors directly into one or more agents, enabling selective triggering while preserving normal outputs in benign cases.

This work fills the gap by investigating backdoor vulnerabilities in multi-agent systems and proposing a defense mechanism that leverages agents’ reasoning abilities and their interactions. Backdoor attacks cause LLM agents to learn a ‘shortcut’ from the trigger to the target output, bypassing logical reasoning. To mitigate this, we design demonstrations that encourage agents to explicitly generate reasoning steps, thereby reducing the likelihood of blindly following attackinduced shortcuts. Agents then inspect each other’s reasoning process to identify inconsistencies between the rationale and the final answer. Any such inconsistency signals a lack of valid support and suggests potential backdoor manipulation. We integrate this defense strategy into existing multi-agent frameworks without disrupting their original interaction flow, thereby enhancing robustness in a plug-and-play manner. In summary, our main contributions are:

· We propose PeerGuard: a collaborative defense strategy for multi-agent systems, in which agents autonomously verify each other’s reasoning to detect backdoor-induced inconsistencies, enhancing overall system trustworthiness.
· We empirically validate the proposed method on diverse benchmarks, demonstrating strong defense performance in GPT- and Llama3-based multi-agent systems.

II. RELATED WORK

LLMMulti-Agent Systems: Challenges and Open Problems

Shanshan Han 1 Qifan Zhang 1 Yuhang Yao 2 Weizhao Jin 3 Zhaozhuo Xu 4

Abstract

This paper explores multi-agent systems and identify challenges that remain inadequately addressed. By leveraging the diverse capabilities and roles of individual agents, multi-agent systems can tackle complex tasks through agent collaboration. We discuss optimizing task allocation, fostering robust reasoning through iterative debates, managing complex and layered context information, and enhancing memory management to support the intricate interactions within multiagent systems. We also explore potential applications of multi-agent systems in blockchain systems to shed light on their future development and application in real-world distributed systems.

1. Introduction

Multi-agent systems enhance the capabilities of single LLM agents by leveraging collaborations among agents and their specialized abilities (Talebirad & Nadiri, 2023; Zhang et al., 2023a; Park et al., 2023; Li et al., 2023; Jinxin et al., 2023). It utilizing collaboration and coordination among agents to execute tasks that are beyond the capability of any individual agent. In multi-agent systems, each agent is equipped with distinctive capabilities and roles, collaborating towards the fulfillment of some common objectives. Such collaboration, characterized by activities such as debate and reflection, has proven particularly effective for tasks requiring deep thought and innovation. Recent works include simulating interactive environments (Park et al., 2023; Jinxin et al., 2023), roleplaying (Li et al., 2023), reasoning (Du et al., 2023; Liang et al., 2023), demonstrating the huge potential of multi-agent systems in handling complex real-world scenarios.

While existing works have demonstrated the impressive capabilities of multi-agent systems, the potential for advanced multi-agent systems far exceeds the progress made to date.

1 University of California, Irvine, CA, USA 2 Carnegie Mellon University, Pittsburgh, PA, USA 3 University of Southern California, Los Angeles, CA, USA 4 Stevens Institute of Technology, Hoboken, NJ, USA. Correspondence to: Shanshan Han < shanshan.han@uci.edu > .

Alarge number of existing works focus on devising planning strategies within a single agent by breaking down the tasks into smaller, more manageable tasks (Chen et al., 2022; Ziqi &Lu, 2023; Yao et al., 2023; Long, 2023; Besta et al., 2023; Wang et al., 2022b). Yet, multi-agent systems involve agents of various specializations and more complex interactions and layered context information, which poses challenges to the designing of the work flow as well as the whole system. Also, existing literature pays limited attention to memory storage, while memory plays a critical role in collaborations between agents. It enables agents to access to some common sense, aligning context with their tasks, and further, learn from past work flows and adapt their strategies accordingly.

To date, multiple significant challenges that differentiate multi-agent systems and single-agent systems remain inadequately addressed. We summarize them as follows.

· Optimizing task allocation to leverage agents’ unique skills and specializations.
· Fostering robust reasoning through iterative debates or discussions among a subset of agents to enhance intermediate results.
· Managing complex and layered context information, such as context for overall tasks, single agents, and some common knowledge between agents, while ensuring alignment to the general objective.
· Managing various types of memory that serve for different objectives in coherent to the interactions in multiagent systems

This paper explores multi-agent systems, offering a survey of the existing works while shedding light on the challenges and open problems in it. We study major components in multi-agent systems, including planning and memory storage, and address unique challenges posed by multiagent systems, compared with single-agent systems. We also explore potential application of multi-agent systems in blockchain systems from two perspectives, including 1) utilizing multi-agent systems as tools, and 2) assigning an agent to each blockchain node to make it represent the user, such that the agent can can complete some tasks on behalf of the user in the blockchain network.

2. Overview

2.1. Structure of Multi-agent Systems

The structure of multi-agent systems can be categorized into various types, based on the each agent’s functionality and their interactions.
Report generation prompt:
Based on the following research information:
— START OF RESEARCH INFORMATION —
Title: VAIN: Attentional Multi-agent Predictive Modeling
Date: 2017-06-19
PDF Url: http://arxiv.org/pdf/1706.06122v2