Comprehensive Survey Maps Five Core Modules Driving Large Language Model Multi-Agent Systems

In a comprehensive systematic survey published in Vicinagearth, researchers led by Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang examined the rapid maturation of autonomous intelligent systems powered by large language models. Historically, intelligent agent architectures were heavily dominated by reinforcement learning frameworks, wherein individual agents carried out constrained tasks with limited environmental feedback and modest adaptability. The emergence of modern foundation models endowed with sophisticated reasoning, contextual understanding, and planning capabilities has catalyzed a decisive paradigm shift toward multi-agent systems. These architectures deploy groups of specialized autonomous agents capable of communicating, coordinating, and solving intricate problems collaboratively. Multi-agent frameworks represent a critical research avenue toward realizing artificial general intelligence, emulating human societal dynamics to harness collective intelligence while preserving individual agent specializations.

To resolve fragmentation across disparate studies, the survey synthesizes existing literature into a cohesive, five-component operational workflow defining the complete multi-agent lifecycle. This cognitive framework categorizes agent execution into five fundamental stages: profile formulation, sensory perception, internal self-action, mutual interaction, and continuous evolution. Rather than viewing agents as monolithic prompt wrappers, the researchers conceptualize them as modular entities that initialize distinct personas, gather environmental stimuli across various modalities, access structured memory records, deliberate through structured cognitive chains, coordinate with peers, and progressively adapt their behaviors. By delineating these interconnected components, the authors offer a unified roadmap encapsulating earlier single-agent and multi-agent milestones. This architectural blueprint establishes a structured basis for analyzing how autonomous software entities interpret dynamic objectives, negotiate complex problem spaces, and systematically execute downstream applications.

The workflow commences with the agent profile module, which instantiates independent intelligent entities endowed with specific identities, occupational roles, psychological attributes, and behavioral constraints. Crafting meticulous profiles ensures that individual agents remain properly calibrated for assigned subtasks, mirroring specialized human roles within institutions or corporate organizations. The authors categorize profile construction strategies into three primary methodologies: contextualized generation, pre-defined assignment, and learning-based generation. Contextualized generation manually decomposes complex operational scenarios into specialized roles, directly embedding customized descriptions, objectives, and behavioral boundaries into initial prompt instructions. Conversely, pre-defined methods leverage foundation models to synthesize extensive pools of diverse personas from which appropriate agents are dynamically selected. Finally, learning-based approaches combine pre-existing seed configurations with automated generative rules, enabling systems to dynamically spawn novel specialized agents during runtime execution as shifting task requirements dictate.

Following profile generation, the perception module enables agents to capture, process, and convert external environmental signals and internal operational states into actionable intermediate representations. While earlier agent frameworks relied almost exclusively on textual descriptions, the integration of multimodal foundation models has broadened sensory acquisition to include visual imagery, continuous video streams, and auditory signals. The survey categorizes sensory inputs into entire environment messages detailing background spatial layouts, interaction messages recording inter-agent dialogue, and self-reflection messages tracking internal cognitive adjustments. To ingest visual data, systems employ visual language models as feature adapters, transform image patches into discrete visual tokens via Vision Transformers, or align visual representations with language embeddings using specialized Querying Transformer interfaces. Auditory perception similarly captures speech rhythms, emotional nuances, and acoustic events by embedding speech tokens into model vocabularies or converting spectrograms into visual patches.

At the center of individual autonomy lies the self-action module, wherein agents process perceived information to retrieve memories, deliberate logically, and execute concrete behavioral responses. The memory sub-module operates as an essential storage and retrieval apparatus, enabling agents to leverage historical experiences and dynamically update cognitive models. During memory retrieval, systems extract relevant experiential records by evaluating quantitative metrics such as recency, topical relevance, and subjective importance, scoring stored items to determine which information should populate prompt contexts. Memory storage mechanisms archive interaction histories and environmental feedback through structured natural language text, dense embedding vectors, triplet configurations, or external relational database queries mediated by structured query language commands. To maintain operational efficiency within finite context windows, systems execute memory reflection routines, distilling raw episodic logs into higher-level abstract insights while pruning obsolete records through first-in-first-out eviction policies or relevance-based deletion.

Beyond internalized memory records, self-action relies heavily on knowledge utilization mechanisms integrating external databases, web scraping interfaces, and specialized retrieval-augmented generation pipelines. Agents access curated external repositories such as knowledge graphs and biomedical archives to supplement pretrained parametric knowledge with verifiable factual data. In retrieval-augmented generation workflows, models retrieve query-relevant document segments, re-rank contextual fragments, and synthesize grounded responses, employing iterative, recursive, or adaptive retrieval strategies to handle multi-step inquiries. However, external knowledge extraction faces persistent hurdles, notably factual hallucinations, temporal knowledge obsolescence, and spurious training biases. Foundation models frequently generate plausible yet unsubstantiated claims, requiring integration with automated fact-checking systems, verification queries, and debiased sampling techniques. Because static model weights cannot easily incorporate newly emerging facts without incurring catastrophic forgetting or prohibitive retraining costs, modular retrieval interfaces remain indispensable for maintaining temporal accuracy.

Cognitive reasoning and deliberate task planning constitute another vital dimension of self-action, determining how agents deconstruct overarching goals into structured operational steps. The survey contrasts one-step reasoning frameworks with iterative multi-step methodologies. In one-step formulations, models decompose tasks into sequential sub-goals through structured elicitation techniques like chain-of-thought prompting, zero-shot directives, and multi-path exploration methods including Tree of Thought and Graph of Thought structures. Conversely, multi-step methodologies engage models in continuous deliberative cycles, dynamically generating incremental action steps while adapting to real-time observations and environmental feedback. Frameworks such as SwiftSage balance rapid intuitive heuristics with deliberative reasoning, while DECKARD separates exploratory planning into distinct dreaming and wakeful execution stages. When domain-specific challenges exceed native language model capabilities, agents integrate with external algorithmic reasoners, translating environmental states into formal Planning Domain Definition Language definitions to compute precise, mathematically verified plans.

Action execution translates formulated plans into tangible behavioral changes within interactive environments, spanning immediate decisions, pre-defined trajectories, and dynamic hybrid creations. Instant decision-making models query memory banks and prompt instructions to trigger rapid actions, whereas pre-defined planning approaches enforce strict adherence to predetermined schedules unless explicit failure signals emerge. Dynamic creation architectures harmonize both paradigms, maintaining broad strategic objectives while granting agents the flexibility to adjust localized behaviors based on real-time feedback. Operationally, agent actions manifest across three major functional domains: task-driven execution, communication interaction, and autonomous environment exploration. In task-driven environments, agents interact with external software tools, code interpreters, and application programming interfaces to automate coding, calculate values, or manipulate digital infrastructure. In exploratory contexts, embodied agents iteratively test novel physical skills and navigate unfamiliar digital simulations, continuously refining operational strategies through empirical trial and error.

The mutual interaction module governs how autonomous agents communicate, organize collective relationships, and coordinate joint activities within shared operational environments. Message delivery mechanisms facilitate information transfer through direct point-to-point dialogues, broadcast signals, or indirect access via centralized shared memory pools where agents publish and subscribe to relevant message streams. Structurally, communication networks organize into hierarchical frameworks where supervisory agents issue binding directives to subordinates, decentralized peer-to-peer topologies eliminating central failure points, and centralized architectures orchestrating interactions through a primary planning model. These structural configurations operate across cooperative, adversarial, and mixed interaction scenarios. Cooperative setups coordinate agents toward shared objectives through standardized operating procedures and consensus protocols. Adversarial scenarios leverage competitive debates between opposing models to sharpen factual consistency and eliminate cognitive bias. Meanwhile, mixed frameworks combine parallel independent task processing with hierarchical supervision to optimize large-scale organizational workflows.

To prevent static operational decay, the evolution module establishes mechanisms for agents to continuously update cognitive structures, refine decision-making strategies, and expand problem-solving capabilities over time. Evolutionary refinement is driven by external feedback collected from surrounding physical or virtual environments, inter-agent communicative exchanges, and human oversight signals aligning outputs with societal norms. Methodologically, systems achieve behavioral evolution through parameter fine-tuning, feedback learning, automated prompt engineering, and reinforcement learning. Fine-tuning approaches range from comprehensive full-parameter updates to parameter-efficient techniques such as Low-Rank Adaptation and prefix tuning, adapting foundation models with minimal computational overhead. Non-parametric alternatives, exemplified by the Reflexion framework, implement verbal feedback learning, storing linguistic post-mortem reflections within episodic memory to guide future operational attempts without altering underlying network weights. Concurrently, agents undergo dynamic architectural adjustments, autonomously modifying memory indexes, revising action plans, or scaling agent counts to address evolving operational bottlenecks.

The authors thoroughly chronicle the extensive deployment of multi-agent architectures across practical problem-solving domains and realistic world simulation environments. In problem-solving spheres, collaborative agent teams automate complex software engineering pipelines, spanning code generation, unit testing, fault reproduction, and hardware register-transfer design. Industrial applications leverage multi-agent digital twins for automated manufacturing control and laser cavity optimization, while embodied robotic platforms utilize foundation models for spatial reasoning, multi-arm path negotiation, and personalized domestic assistance. In laboratory sciences, specialized agent platforms autonomously plan organic chemical syntheses, coordinate physical experimental apparatus, and design de novo functional proteins. Within world simulation settings, multi-agent frameworks emulate non-player gaming behaviors, generate procedural environments, simulate microeconomic market negotiations under varying social preferences, and model interpersonal dynamics in virtual towns. Furthermore, epidemiological researchers deploy generative agent populations to simulate behavioral responses to disease outbreaks, such as self-quarantining protocols, yielding authentic contagion curves.

Despite notable achievements, the survey underscores formidable open challenges currently hindering the dependable real-world deployment of large language model multi-agent systems. Foundation models remain vulnerable to black-box opacity, severe factual hallucinations, and systematic training biases that degrade decision reliability in high-stakes environments. Furthermore, malicious exploitation risks loom large, including automated cyber-attacks, synthetic disinformation campaigns, and adversarial prompt injections, necessitating robust defense filtering and stringent ethical governance frameworks. Scaling multi-agent systems also introduces severe computational bottlenecks, communication latency, and coordination friction, driving demand for dynamic agent allocation, token-efficient publish-subscribe channels, and tiered model architectures. Looking forward, the researchers identify collective intelligence, agent-as-a-service cloud platforms, and cross-domain multimodal integration as pivotal frontiers. Perfecting coordinated group reasoning and democratizing access through scalable infrastructure will be essential to transforming experimental multi-agent prototypes into robust, trustworthy, and ubiquitously deployed autonomous services.

Subject of Research: Architectural frameworks, operational workflows, and challenges of large language model multi-agent systems

Article Title: A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges

Article References: Li, X., Wang, S., Zeng, S., Wu, Y., & Yang, Y. (2024). A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth, 1(1), Article 9. https://doi.org/10.1007/s44336-024-00009-2

Image Credits: AI Generated

DOI: 10.1007/s44336-024-00009-2

Keywords: multi-agent systems, large language models, autonomous agents, agent profiles, multimodal perception, self-action mechanisms, memory retrieval, retrieval-augmented generation, mutual interaction, agent evolution, cooperative systems, adversarial interaction, world simulation, collective intelligence, hallucination mitigation, agent as a service, reinforcement learning