PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems

Bingyu Yan, Zhibo Zhou, Litian Zhang, Lian Zhang, Ziyi Zhou, Dezhuang Miao, Zhoujun Li, Chaozhuo Li, Xiaoming Zhang

arXiv:2502.14321v3cs.MAcs.CL

TL;DR

Existing LLM-MAS surveys often under-cover communication and coordination workflows, despite communication’s central role in multi-agent collaboration. This survey develops a communication-centric framework spanning system-level and internal communication, synthesizes recent literature, and identifies interoperability, multimodal integration, security, benchmarking, and scalability challenges.

  • Problem

    Existing surveys often categorize LLM-MAS by applications or architectures while providing limited coverage of communication and coordination workflows.

  • Method

    The survey analyzes LLM-MAS through system-level communication and system-internal communication, covering architectures, goals, protocols, strategies, paradigms, objects, and content.

  • Results

    The survey systematically explores communication components and summarizes their impacts on LLM-MAS efficiency, scalability, and adaptability.

  • Takeaways & Limitations

    The review identifies communication mechanisms and future directions concerning unified protocols, multimodal integration, security, system design, and evaluation.

  • Takeaways & Limitations

    LLM-MAS face interoperability barriers because agent groups using distinct protocols cannot seamlessly communicate or collaborate.

Abstract

from arXiv · show

Large language model-based multi-agent systems have recently gained significant attention due to their potential for complex, collaborative, and intelligent problem-solving capabilities. Existing surveys typically categorize LLM-based multi-agent systems (LLM-MAS) according to their application domains or architectures, overlooking the central role of communication in coordinating agent behaviors and interactions. To address this gap, this paper presents a comprehensive survey of LLM-MAS from a communication-centric perspective. Specifically, we propose a structured framework that integrates system-level communication (architecture, goals, and protocols) with system internal communication (strategies, paradigms, objects, and content), enabling a detailed exploration of how agents interact, negotiate, and achieve collective intelligence. Through an extensive analysis of recent literature, we identify key components in multiple dimensions and summarize their strengths and limitations. In addition, we highlight current challenges, including communication efficiency, security vulnerabilities, inadequate benchmarking, and scalability issues, and outline promising future research directions. This review aims to help researchers and practitioners gain a clear understanding of the communication mechanisms in LLM-MAS, thereby facilitating the design and deployment of robust, scalable, and secure multi-agent systems.

2.1 LLM-Based Agents

LLM-based agents use an LLM as their core reasoning component, while LLM-MAS combine multiple agents with communication and established MAS coordination principles. The survey organizes these systems through communication architectures and workflows.

  • LLM-based agents are autonomous entities whose core reasoning component is an LLM.
  • Each agent typically combines Brain, Perception, and Action components for reasoning, multimodal input processing, and tool-integrated execution.The Brain integrates short- and long-term memory; Perception converts sensory inputs into prompts, and Actions include external tools or actuators.
  • Traditional MAS contribute symbolic communication languages, protocol templates, and learning-based coordination methods, but often face brittleness, task-specific messaging, sample complexity, or limited interpretability.The supplied passages describe machine-interpretable speech acts, contract nets, auctions, centralized training, graph messages, and related limitations.
  • LLM-MAS combine LLM linguistic flexibility with MAS coordination, allowing agents to cooperate, compete, or negotiate according to system objectives.
  • The survey analyzes LLM-MAS communication through architectures, goals, protocols, and internal communication workflows.Its system-level perspective examines how agents are organized and communicate, while the broader structure includes foundational concepts and communication-centric analysis.

3.1.1 Flat Architecture

Flat architecture uses decentralized peer-to-peer interactions without hierarchy or central oversight, supporting agile coordination but creating scalability challenges as agent populations grow.

  • Flat architecture organizes agents as decentralized peers without hierarchical distinctions or central oversight.
  • This architecture suits agile, flexible, and spontaneous interactions such as dynamic task assignment and rapid decision-making.
  • Increasing the number of agents can raise communication overhead and reduce coordination efficiency in flat systems.
  • Alternatives: Hierarchical and team-based architectures instead emphasize layered delegation or specialized group collaboration, illustrating alternatives to flat organization.

3.1.4 Society Architecture

Society architectures place diverse agents in broader social environments governed by shared norms or rules, enabling large-scale simulations of emergent behavior. The survey also contrasts this with specialized and mixed communication structures and goals.

  • Society architectures model diverse agents, roles, motivations, and interactions within broader social environments governed by shared norms or rules.
  • Generative Agents and EconAgent demonstrate society architectures through friendship formation, activity organization, adaptive behavior, and simulated economic decisions.
  • Hybrid architectures combine communication structures, including hierarchical oversight and decentralized collaboration, to adapt interactions and resource allocation.
  • Architecture choices influence capabilities, efficiency, and scalability, requiring trade-offs among flexibility, control, scalability, and efficiency.
  • Communication goals: Communication goals include cooperation, competition, and mixed goals, with cooperation further divided into direct cooperation and cooperation through debate.
  • Communication goals: Direct cooperation shares information and resources, whereas debate-based cooperation critiques and refines inputs to support more robust conclusions.

3.2.2 Competition

Competitive LLM-MAS interactions arise from conflicting objectives or scarce resources and can stimulate strategic and adaptive behavior. However, adversarial or deceptive behavior creates instability and security risks requiring regulatory constraints.

  • Competitive interactions occur when agents have conflicting objectives or compete for limited resources or favorable outcomes.
  • Adversarial settings can stimulate strategic thinking, innovation, and adaptive behavior as agents refine their strategies.
  • Competitive agents may shape interactions through persuasive or deceptive narratives, including in game-playing, language-evolution, and competitive-debate settings.
  • Adversarial or deceptive behaviors can induce instability and security vulnerabilities in competitive systems.
  • Sustainable and robust performance requires reconciling competitive incentives with cooperative regulatory constraints.

3.3.1 Model Context Protocol

Emerging protocols standardize how LLM-MAS agents interact with external resources and with one another, addressing interoperability, security, and scalable coordination. MCP uses a structured client-server model, while A2A and ANP support peer and decentralized communication.

  • Model Context Protocol: MCP enables secure, structured interactions between LLM agents and external tools, data, and services through a JSON-RPC client-server architecture.Hosts request context, clients manage connections, and servers execute requests against resources.
  • Model Context Protocol: MCP improves interoperability and scalability by reducing fragmentation across agents and tool providers.Its layered architecture also decouples sensitive tool invocations from LLM-generated responses to reduce data-exposure risks.
  • Emerging protocols: A2A supports secure peer-to-peer agent communication through capability-based Agent Cards that enable dynamic task delegation.Agents advertise capabilities over HTTP and Server-Sent Events for enterprise-scale collaboration.
  • Emerging protocols: ANP enables decentralized discovery and communication across open networks using decentralized identifiers, JSON-LD graphs, encryption, and secure channels.These mechanisms support interoperable interactions among heterogeneous agents across network boundaries.
  • Emerging protocols: MCP, A2A, and ANP exemplify an evolving protocol landscape that also includes ACP, AITP, and AConP.The comparative framework adds dimensions such as security and use cases to contextualize these protocols for LLM-MAS.

4.1.1 One-by-One

Communication strategies determine when and in what sequence LLM-MAS agents exchange information. Sequential interaction supports coherent integration, whereas concurrent interaction increases speed and diversity but introduces synchronization and conflict challenges.

  • One-by-One: One-by-One communication makes agents respond sequentially after processing previous messages, preserving context and reducing misunderstandings.It is particularly suited to methodical reasoning and detailed integration of multiple agents’ inputs.
  • One-by-One: Chain-of-Agents applies One-by-One communication by having each agent summarize prior contributions before adding input to lengthy text-generation tasks.This helps manage contexts beyond typical token limits.
  • One-by-One: Latency scales linearly with the number of turns in One-by-One communication, while early inaccuracies can propagate and cause semantic drift.Each agent waits for preceding exchanges to finish before contributing.
  • Simultaneous-Talk: Simultaneous-Talk enables concurrent communication, supporting rapid idea generation, parallel problem-solving, and diverse perspectives.Autoagents uses concurrent solution-path generation, while EconAgent applies a similar pattern to macroeconomic simulations.
  • Simultaneous-Talk: Simultaneous-Talk can produce state staleness, redundant or inconsistent outputs, and conflicting proposals that require arbitration.Effective timestamps and lightweight consensus mechanisms are identified as important safeguards.

4.1.3 Simultaneous-Talk-with-Summarizer

Simultaneous-Talk with Summarizer combines concurrent agent exchanges with periodic consolidation by a summarizing agent. This improves shared alignment but retains sequential dependencies and creates risks when summaries misrepresent agent communications.

  • Simultaneous-Talk-with-Summarizer: Simultaneous-Talk with Summarizer adds a summarizing agent that consolidates concurrent communications into coherent, unified messages.The summarizer periodically synthesizes inputs so agents remain aligned with the overall system context.
  • Survey classification: Table 3 classifies LLM-MAS studies by communication strategy, paradigm, object, and content.The listed strategies are One-by-one, Simultaneous-Talk, and Simultaneous-Talk-with-Summarizer; paradigms include message passing, speech act, and blackboard.
  • Simultaneous-Talk-with-Summarizer: Summarizers can aggregate communication outcomes in hierarchical and team-based coordination scenarios to support decision-making and task delegation.CausalGPT and AgentCoord are cited as examples.
  • Simultaneous-Talk-with-Summarizer: Adding a summarizer mitigates synchronization issues but reintroduces sequential dependencies into the communication process.The strategy therefore combines parallel interaction with an additional consolidation stage.
  • Simultaneous-Talk-with-Summarizer: Inaccurate summarization can misrepresent agent communications or introduce hallucinations that compromise system effectiveness.The summarizer must correctly interpret and compress the agents’ messages.

4.2.2 Speech Act

The Speech Act paradigm treats communication as performative action: agent utterances are designed to influence behaviors or trigger system state changes. Its expressive coordination benefits come with risks from ambiguity, underspecified context, and divergent beliefs.

  • Speech Act: Speech Act communication uses instructive, persuasive, or directive utterances to trigger agent actions or system state changes.The paradigm treats language as an action-performing mechanism rather than only an information channel.
  • Speech Act: Speech acts support diplomatic negotiation, collaborative reasoning, debate-driven optimization, and software co-engineering workflows.Directive, commissive, and persuasive components can coordinate task allocation, belief alignment, and temporal commitments.
  • Speech Act: Ambiguous force marking, incomplete contextual frames, and divergent beliefs can cause unintended goals, deadlocks, or cascading errors.The recipient may act on incomplete preconditions or infer a meaning different from the sender’s intent.
  • Blackboard: Blackboard communication provides a centralized repository where agents share, retrieve, and coordinate through published messages or updates.It supports a unified understanding of system states, tasks, and decisions in highly coordinated settings.
  • Blackboard: Blackboard systems require safeguards against bottlenecks, unauthorized access, misinformation, and malicious inputs.These risks arise alongside their benefits for information consistency and accessibility.

4.3.1 Communication with Self

Communication with self describes internal dialogue that lets agents deliberate, plan, refine decisions, and update internal states. The paper also distinguishes communication with other agents and the environment, each introducing coordination or perception requirements.

  • Communication with Self: Communication with self uses internal dialogue or reflection for independent deliberation, planning, decision refinement, and internal-state updates.AgentCoord and FixAgent illustrate iterative strategy evaluation and debugging-oriented self-reflection.
  • Communication with Self: Internal communication requires robust cognitive frameworks to simulate reflection accurately and manage agent cognition.
  • Communication with Other Agents: Communication with other agents covers information exchange, task coordination, and negotiation for collaborative problem-solving and coordinated decisions.MAGIS and AgentFM exemplify structured inter-agent exchanges in software development and database failure management.
  • Communication with Other Agents: Effective inter-agent communication requires protocols that support synchronization, consistency, and conflict resolution, especially in large-scale systems.
  • Communication with the Environment: Communication with the environment converts multimodal sensor streams or simulator state into actionable knowledge while agents adapt to real-time external changes.EmbodiedGPT processes visual and audio inputs, while ChatSim incorporates environmental feedback during autonomous-driving simulations.
  • Communication with the Environment: Environment communication demands low-latency sensor fusion, continual grounding in non-linguistic data, and safeguards against perception errors or out-of-distribution inputs.These requirements are identified as important for dynamic, high-stakes deployments.

4.3.4 Communication with Human

Communication with humans adds an interaction layer in which agents interpret and respond to user inputs, commands, and feedback. Applications in education and medicine illustrate this modality, alongside challenges involving understanding, rapport, safety, and ethics.

  • Communication with Human: Human-agent communication requires agents to interpret and appropriately respond to human inputs, commands, or feedback during natural interactions.
  • Communication with Human: PeerGPT applies human-agent communication in collaborative education, where agents interact with children as moderators or participants.Agents interpret verbal inputs and physical actions while refining responses to support learning and engagement.
  • Communication with Human: Human-agent deployment faces challenges in natural-language understanding, affective and situational signal integration, privacy, bias, and accountability.
  • Communication Objects: Clearly managing communication objects influences interaction adaptability, responsiveness, and overall operational effectiveness in LLM-MAS.

4.4.1 Explicit communication

Explicit communication exchanges information directly through interpretable natural language or precise code and structured data. These forms support flexible contextual interaction or accurate standardized exchanges, respectively.

  • Explicit Communication: Explicit communication conveys information directly with clearly defined and easily interpretable meaning.
  • Natural Language: Natural-language communication enables agents to exchange clear, flexible, and contextual messages in human-readable text or speech.Its flexibility and expressiveness support negotiation, planning, conflict resolution, and strategic coordination.
  • Code and Structured Data: Code and structured data provide precise, unambiguous exchanges suited to instructions, algorithmic descriptions, and structured information transfer.
  • Code and Structured Data: Structured communication can reduce misinterpretations but requires strict adherence to predefined data standards and formats.MAGIS and AutoData apply structured exchanges to issue tracking, code management, and open-web data collection.

4.4.2 Implicit communication.

Implicit communication conveys information through agent behavior or environmental conditions rather than explicit statements. The paper connects these mechanisms to adaptive interaction while identifying broader system challenges involving interpretation, architecture, protocols, multimodality, security, and evaluation.

  • Implicit Communication: Implicit communication conveys information indirectly through actions, environmental cues, contextual interpretation, or feedback.
  • Behavioral Feedback: Behavioral feedback communicates intentions, commitments, private information, or strategic adjustments through observable actions and adaptive responses.Richelieu and Flooding simulations illustrate behavioral signaling in negotiation and peer adaptation.
  • Challenges: Behavioral feedback depends on accurate interpretation mechanisms, while larger systems require efficient and scalable communication paradigms.
  • Environmental Signal: Environmental signals influence agent decisions when agents interpret changing operational conditions and adapt their actions.ChatSim uses traffic and weather conditions, while EcoLANG uses economic and environmental indicators.
  • Communication Protocols: Protocol proliferation creates functional redundancy and management complexity, while poor interoperability hinders communication, collaboration, scalability, and integration.
  • Communication Protocols: A unified standardized protocol is proposed to improve interoperability, security, integration, and management simplicity across disparate LLM-MAS.
  • Multimodal Communication: Multimodal communication should extend beyond text to images, audio, and video, but coherent cross-modal coordination and processing remain challenges.
  • Security: Safety-critical LLM-MAS require communication-layer protection against eavesdropping, forgery, man-in-the-middle attacks, data tampering, and spoofing.
Loading 2502.14321v3…