PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Agent Collaboration Mechanisms: A Survey of LLMs

Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O'Sullivan, Hoang D. Nguyen

arXiv:2501.06322v1cs.AI

TL;DR

LLM-based MAS research needs a fuller account of how agents collaborate toward shared goals, beyond the limitations and isolated capabilities of individual LLMs. This paper surveys collaboration mechanisms and organizes them in an extensible framework, finding that effective channel design is central while evaluation, safety, and complexity remain important boundaries.

  • Problem

    Existing research does not fully explore the collaborative aspects and mechanisms needed for LLM-based agents to work together toward shared goals.

  • Method

    The paper systematically surveys MAS research and characterizes collaboration through actors, types, structures, strategies, and coordination mechanisms.

  • Results

    The survey synthesizes collaboration findings into a structured, extensible framework for analyzing and designing collaborative interactions among LLM-based agents.

  • Takeaways & Limitations

    Effective collaboration channels and clear protocols are crucial, while poorly designed competitive channels can be overtaken by a strong single-agent system.

  • Takeaways & Limitations

    MAS evaluations often use narrow, inconsistent scenarios without standardized protocols, limiting objective comparison and progress tracking.

Abstract

from arXiv · show

With recent advances in Large Language Models (LLMs), Agentic AI has become phenomenal in real-world applications, moving toward multiple LLM-based agents to perceive, learn, reason, and act collaboratively. These LLM-based Multi-Agent Systems (MASs) enable groups of intelligent agents to coordinate and solve complex tasks collectively at scale, transitioning from isolated models to collaboration-centric approaches. This work provides an extensive survey of the collaborative aspect of MASs and introduces an extensible framework to guide future research. Our framework characterizes collaboration mechanisms based on key dimensions: actors (agents involved), types (e.g., cooperation, competition, or coopetition), structures (e.g., peer-to-peer, centralized, or distributed), strategies (e.g., role-based or model-based), and coordination protocols. Through a review of existing methodologies, our findings serve as a foundation for demystifying and advancing LLM-based MASs toward more intelligent and collaborative solutions for complex, real-world use cases. In addition, various applications of MASs across diverse domains, including 5G/6G networks, Industry 5.0, question answering, and social and cultural settings, are also investigated, demonstrating their wider adoption and broader impacts. Finally, we identify key lessons learned, open challenges, and potential research directions of MASs towards artificial collective intelligence.

1 Introduction

LLM-based MASs address limitations of isolated LLMs by coordinating specialized agents, but prior surveys insufficiently examine collaboration mechanisms. This survey develops a comprehensive framework, reviews applications, and identifies challenges and research directions.

  • LLM-based MASs use collaboration to help agents connect, negotiate, decide, plan, and act jointly rather than operate as isolated models.
  • Prior surveys often focus on single-agent systems or treat multi-agent collaboration superficially, leaving collaboration mechanisms insufficiently explored.
  • The survey characterizes collaboration through actors, types, structures, strategies, and coordination architecture, emphasizing operational know-how for effective collaboration.
  • It reviews real-world LLM-based MAS implementations across diverse domains and discusses their practical applications, successes, and limitations.
  • The paper identifies challenges including collective reasoning and decision-making and outlines potential research directions.

2 Background

Multi-agent systems combine interacting intelligent agents, environments, and communication to solve complex tasks through distributed collaboration. Collaborative AI extends this setting to AI-agent or human cooperation, while LLMs provide agents with language and reasoning capabilities alongside important limitations.

  • 2.1 Multi-Agent (AI) Systems: A multi-agent system is a computerized system of multiple interacting intelligent agents situated in an environment.
  • 2.1 Multi-Agent (AI) Systems: MASs divide complex problems among agents that learn, reason, make autonomous decisions, and coordinate through interactions with neighboring agents or environments.
  • 2.1 Multi-Agent (AI) Systems: MASs offer flexibility, scalability, robustness, reliability, self-organization, and coordination through distributed operation and agent-level task allocation.
  • 2.2 Large Language Models: LLMs provide agents with broad language understanding, generation, reasoning, and emergent abilities, but face changing knowledge, computational, ethical, and misuse-related challenges.
  • 2.3 Collaborative AI: Collaborative AI concerns systems working with other AI agents or humans, and MAS research studies agents that learn, adapt, and collaborate toward shared goals.
  • 2.3 Collaborative AI: Collaboration spans cooperation, competition, and coopetition, with LLM-based MAS research emphasizing mechanisms such as debate, negotiation, and competition.

3 Multi-Agent Collaboration Concept

The paper formalizes agents and multi-agent systems around models, objectives, environments, inputs, outputs, and shared goals. It then treats collaboration channels as structured mechanisms for coordinating agents across communication stages and interaction types.

  • 3.1 Agent and Collaborative System Definition: An agent is represented as a = {m, o, e, x, y}, comprising a model, objective, environment, input perception, and output or action.
  • 3.1 Agent and Collaborative System Definition: The agent model includes architecture, specific memory, and optional adapters, while its objective guides actions within the system.
  • 3.1 Agent and Collaborative System Definition: An agent produces output y = m(o, e, x) by applying its model to an objective, environment, and input.
  • 3.1 Agent and Collaborative System Definition: A multi-agent system contains agents, collective goals, and a shared environment, with agent numbers potentially predefined or dynamically adjusted.
  • 3.2 Collaboration Mechanisms: Collaboration occurs at late stages through output ensembling, mid stages through model exchange, and early stages through data, context, or environment sharing.
  • 3.2 Collaboration Mechanisms: Collaboration channels coordinate agents through communication and deeper management behaviors, producing collective actions from agents’ objectives, environments, and inputs.
  • 3.2 Collaboration Mechanisms: Channels differ by actors, type, structure, and strategy, including cooperation, competition, coopetition, peer-to-peer, centralized, and distributed arrangements.

4.1 Overview

The survey systematically analyzes existing MAS research and synthesizes its findings into a framework centered on collaboration channels. These channels organize coordination by participating agents, interaction type, structure, strategy, and learning-process level.

  • 4.1 Overview: The framework centers on collaboration channels that facilitate coordination and orchestration among agents.
  • 4.1 Overview: The survey systematically analyzes existing MAS research to identify defining collaboration characteristics, implementation trends, and methodologies.
  • 4.1 Overview: Each LLM-based agent is modeled with a language model, objective, environment, input perception, and output or action.
  • 4.1 Overview: Channels are characterized by actors, cooperation or competition type, peer-to-peer or centralized structure, and role-based, rule-based, or model-based strategy.
  • 4.1 Overview: Collaboration mechanisms span data exchange, shared input embeddings, model sharing, and output sharing across machine-learning processes.

4.2 Collaboration Types

LLM-based MASs organize collaboration as cooperation, competition, coopetition, or hybrid combinations across distinct channels. These types balance complementary expertise, rivalry, negotiation, and coordination complexity when pursuing collective goals.

  • Cooperation: Cooperation aligns agents’ individual objectives with a shared collective goal, using complementary expertise and task specialization.Specialized roles can improve efficiency and reduce completion time for complex problem-solving and collective decision-making.
  • Cooperation: Cooperative systems use feedback, evaluation, reflection, role-based conversations, and flexible communication to improve shared task performance.Examples include Actor–Evaluator feedback loops, Theory of Mind representations, CAMEL, and AutoGen.
  • Competition: Competition arises from conflicting objectives or limited resources, with agents pursuing individual goals through rivalry such as debate or strategic gameplay.Competitive interaction can still be oriented toward a collective goal, as in debate judged by another agent.
  • Competition: Competition can strengthen reasoning, creativity, robustness, and strategic adaptability, but poorly designed competitive channels may be surpassed by strong single-agent prompts.Excessive competition can also hamper alignment when cooperation is desired.
  • Coopetition: Coopetition combines cooperation and competition, allowing agents to pursue shared objectives while competing on other aspects through trade-offs or selective expert competition.Negotiation scenarios and mixture-of-experts models illustrate this blended interaction pattern.
  • Hybrid collaboration: Hybrid MASs assign different interaction types to multiple collaboration channels, enriching coordination while introducing additional management complexity.LEGO uses cooperation for information augmentation followed by competition between Explainer and Critic agents; debate systems combine competing debaters with a cooperating judge.

4.3 Collaboration Strategies

LLM-based MASs use rule-based, role-based, and model-based strategies to structure agent interactions. These strategies trade predictable coordination and modular specialization against adaptability, uncertainty handling, and computational complexity.

  • Rule-based Protocols: Rule-based protocols strictly control agent interactions through predefined rules governing acceptable collaborative inputs and actions.They support predictable, traceable coordination for structured collaboration channels.
  • Rule-based Protocols: Rule-based strategies provide efficiency, predictability, straightforward debugging, and fairness for well-defined procedures, but adapt poorly to unexpected situations.Increasing task complexity can cause the rule set to grow exponentially, complicating scalability and maintenance.
  • Role-based Protocols: Role-based protocols divide the collective objective into segmented responsibilities, assigning agents specialized functions based on domain knowledge.AgentVerse, MetaGPT, RoCo, and BabyAGI illustrate role assignment for alignment, verification, physical interaction, and task prioritization.
  • Role-based Protocols: Role-based techniques improve efficiency, structure, modularity, and module reusability through explicit division of labor.They are suited to environments with well-defined specialized jobs.
  • Model-based Protocols: Model-based protocols use probabilistic decision-making over collaborative inputs, environmental data, and shared goals to anticipate likely outcomes under uncertainty.Theory of Mind, probabilistic graphical models, and probabilistic timed automata support coordination in partially observable or unpredictable environments.
  • Model-based Protocols: Model-based methods provide flexibility and robustness in dynamic contexts, but require complex environment and interaction models and can be computationally costly in real time.Their probabilistic procedures may restrict deployment where low-latency operation is required.

4.4 Communication Structures

Communication structures organize how agents exchange information and distribute control across centralized, decentralized, and hierarchical topologies. These structures support different coordination patterns, from hub-based management to peer-to-peer interaction and layered collaboration.

  • Communication structures: Communication structures are categorized into centralized, decentralized and distributed, and hierarchical topologies.Table 4 summarizes their definitions, advantages, disadvantages, and example scenarios.
  • Centralized structure: Centralized systems connect every agent to a central hub that manages, controls, and coordinates interactions.The central agent may host the LLM, or distributed agents may retain their own LLMs while communicating through the hub.
  • Decentralized structure: Decentralized systems distribute control across agents, which operate using local information and possibly limited communication.Peer-to-peer communication is commonly used in world simulation, debate, and dynamic directed acyclic graph structures.
  • Structure and specialization: Communication structures can support specialized collaboration by assigning agents distinct expertise, roles, or adaptive behaviors.Examples include medical specialists, software-development roles, and agents that infer teammates’ intentions while limiting communication.
  • Hierarchical structure: Hierarchical systems organize agents into layers, with agents mainly interacting within their layer or with adjacent layers.DyLAN adds dynamic agent selection and early stopping by deactivating low-performing agents before final task solving.

4.5 Coordination and Orchestration

Coordination and orchestration govern how multiple collaboration channels are created, ordered, and characterized. The survey distinguishes static architectures based on predefined rules from dynamic architectures that adapt roles and channels to inputs and changing requirements.

  • Static architecture: Static coordination uses domain knowledge and predefined rules to establish collaboration channels.Sequential chaining connects specialist agents so each output feeds the next, with a final summarizer producing the response.
  • Static architecture: Sequential collaboration channels can outperform single-agent chain-of-thought methods on complex tasks such as college-level science multiple-choice questions.The cited setup uses two domain experts followed by a summarizer.
  • Static architecture: Static coordination can encode domain workflows, as in MapCoder’s recall, planning, code-generation, and debugging roles.Predefined channels structure information exchange between specialized agents, including direct planning-to-coding communication.
  • Dynamic architecture: Dynamic coordination assigns roles and defines collaboration channels in real time as environments and task requirements change.Solo Performance Prompting dynamically identifies relevant personas from the input.
  • Dynamic architecture: Graph-based orchestration dynamically constructs a DAG whose nodes are tasks and whose edges encode dependencies and collaboration channels.Agents can execute tasks in parallel or sequence, while a Delegator consolidates completed results into the final response.

4.6 Summary and Lessons Learned

The survey presents effective collaboration as dependent on clear channels, domain knowledge, adaptive assignments, and suitable strategies. It also emphasizes that multi-agent systems face safety, cost, scalability, efficiency, and coordination challenges.

  • Challenges: LLM-based MASs require attention to safety, performance, cost, scalability, and efficiency, especially in competitive scenarios.Reported failures include exploitation and hallucination.
  • Evaluation: Current benchmarks evaluate multi-agent systems using success rate, task outcomes, cost-effectiveness, and collaborative efficiency.These metrics are presented as tools for assessing and improving system design.
  • Lessons learned: Robust collaboration channels and clear protocols are crucial for efficient information exchange among agents.The survey reports that well-designed mechanisms can outperform single-agent systems, whereas suboptimal competitive channels can be overtaken by strong single-agent prompts.
  • Lessons learned: Domain-specific knowledge is important for designing collaboration architectures and effective system prompts.Collaboration channels are often predefined to align agent interactions with domain requirements.
  • Lessons learned: Dynamic role and channel assignment can improve flexibility and performance when agent strengths and task requirements vary.Adaptive assignment allows systems to respond to changing environments and objectives.

5 Applications

The survey reviews LLM-based MAS applications in wireless networks, Industry 5.0, QA/NLG, and social and cultural domains. Across these areas, specialized agents and coordination frameworks support communication, evaluation, data generation, and social simulation.

  • Application domains: Applications span 5G/B6G and Industry 5.0, QA/NLG, and social and cultural domains.Table 6 summarizes represented systems by contributions, advantages, and disadvantages.
  • 5G/B6G and Industry 5.0: LLM-SC models transmitted language sequences to balance semantic-level and technical-level performance in semantic communication.Other wireless applications combine LLMs with RIS for resource allocation and signal decoding in the Internet of Vehicles.
  • 5G/B6G and Industry 5.0: Industry 5.0 systems use local open-source LLM deployments, prompt management, post-processing, and task-specific prompt databases to address privacy, security, and scalability concerns.These components also support capability improvements through integrated prompting methods.
  • QA/NLG: Swarm coordinates specialized agents through routines and handoffs, enabling lightweight transitions for scalable customer-service applications.Its illustrated use case covers sales and refunds.
  • QA/NLG: Self-evolving benchmarks use role-based agents to modify contexts or questions and create more challenging mathematical, logical, and commonsense reasoning instances.The framework includes pre-filtering, creation and verification, and candidate-option formulation roles.
  • QA/NLG: Orca-AgentInstruct uses distinct agentic flows to generate tailored synthetic datasets from raw data sources through a generative teaching approach.The approach targets improved model capabilities in different areas.
  • QA/NLG: QA/NLG applications report improved task capabilities, higher-confidence response evaluation, and higher-quality synthetic training data through multi-agent collaboration.Frameworks such as Swarm, Magentic-One, Bee, and LangChain provide practical mechanisms for orchestrating specialized agents.
  • Social and cultural domains: Social and cultural MAS research examines human behavior, social dynamics, and cultural interactions, including theory of mind and non-verbal action inference.These systems are also discussed as methods for survey research, online experiments, automated content analysis, and agent-based modeling.

6 Open Problems & Discussion

LLM-based MASs face open problems in governance, collective decision-making, reliability, scalability, emergent generalization, and evaluation. Progress requires mechanisms for coordinated roles and decisions, safeguards against propagated errors and deception, scalable architectures, and standardized dynamic benchmarks.

  • Governance and Decision-Making: Collective intelligence requires unified governance for coordination, planning, agent selection, task distribution, role assignment, and adaptation.The section identifies role specialization and adaptive role assignment as important governance challenges.
  • Governance and Decision-Making: Current MAS decision-making methods, including dictatorial control and popular voting, may inadequately represent agent preferences and overconfidence.The paper calls for approaches that improve the diversity and fairness of collective decisions.
  • Reliability and Safety: Hallucinations, overconfidence, misunderstandings, and adversarial attacks can propagate through agent interactions and amplify inaccuracies into cascading effects.The risk arises because agents may reinforce inaccurate outputs during collaboration.
  • Scalability and Emergence: Growing agent populations complicate memory, processing time, coordination channels, interaction efficiency, and bottleneck prevention.Understanding MAS scaling laws is presented as necessary for architectures supporting large-scale collaboration.
  • Scalability and Emergence: Unexpected generalization and emergent collective behaviors remain difficult to identify, foster, and explain across unseen domains.The open problem is understanding how coordinated problem-solving and innovation arise from agent interactions.
  • Evaluation and Benchmarking: MAS evaluation remains underdeveloped, inconsistent across narrow configurations, and constrained by the absence of standardized, dynamic benchmarking protocols.Evaluation must cover system performance, coordination efficiency, contextual appropriateness, and fine-grained agent and collaboration behavior.
  • Reliability and Safety: Human-like MAS behavior can blur artificial and human behavior, potentially fostering user over-reliance on system capabilities.The section connects simulated social and psychological traits with concerns about deception and misplaced trust.

7 Conclusion

The survey introduces a structured, extensible framework for understanding and designing collaboration in LLM-based MASs. It organizes collaboration across five dimensions to guide future research toward more intelligent collaborative systems.

  • Conclusion: The framework characterizes LLM-based MAS collaboration through actors, types, structures, strategies, and coordination mechanisms.These dimensions provide a systematic lens for analyzing and designing collaborative interactions.
Loading 2501.06322v1…