From Classic Multi-Agent Paradigms to Future Large-Foundation-Model-Driven Systems

Multi-agent systems (MASs) have become a core paradigm in AI research, finding applications in robotics [1], social intelligence [2], satellite systems [3] and more. Inspired by biological collectives and complex distributed systems [4][5], MASs study how autonomous agents interact to achieve global cooperation or collective intelligence [6][7][8].

In this review, MASs that do not incorporate large foundation models (LFMs) are referred to as classic multi-agent systems (CMASs) [9][10]. CMASs rely on explicitly designed system models or task‑specific learning mechanisms. From a methodological perspective, existing CMAS research can be divided into model‑based and learning‑based approaches. Model‑based work has established classic problem domains and theoretical frameworks such as consensus control [11], formation control [12], task scheduling [13] and bio‑inspired optimization [14], assuming a well‑modeled system with clear objectives to guarantee stability and performance [9]. However, in environments that are unmodelable, have unknown dynamics, or are partially observable, these assumptions limit applicability, prompting the rise of learning‑based methods such as multi‑agent reinforcement learning (MARL) that enable agents to acquire collaborative strategies without precise models [15]. Despite alleviating model dependence, MARL still suffers from sample inefficiency, stability, interpretability and generalization issues [16].

The limitations of CMASs have motivated the exploration of more general methods with reasoning capabilities, leading to the integration of LFMs with MASs [17]. Within the MAS context, LFMs serve as the cognitive core of agents, allowing them to parse unstructured multimodal inputs, maintain contextual understanding, perform complex reasoning, and generate high‑level actions or interaction messages [18]. This shifts agent operation from predefined system models, hand‑crafted rules, or task‑specific policies toward semantic‑level perception and language‑based interaction, enabling more flexible collaboration [19]. Leveraging the pre‑trained knowledge and reasoning power of LFMs, these systems can execute step‑wise planning, knowledge retrieval and high‑level decision making [20][21]. As illustrated in Figure 1, unlike CMASs that are tailored to fixed environments, LFM‑driven MASs (LMASs) exhibit strong generalization, can accumulate experience across tasks, and support flexible cooperation in open, dynamic scenarios [22][23].

Existing surveys of LMASs mainly focus on LFM‑centric paradigms and summarize system architectures, collaboration mechanisms and application scenarios [8][20][24][19][25]. In contrast, this paper proposes a unified perspective that connects CMASs and LMASs. LMASs are not replacements for CMASs but complementary extensions that enhance classic systems with high‑level reasoning and generalization, while CMASs remain indispensable for reliable low‑level control and theoretical guarantees.

The paper’s contributions are threefold: (1) a comprehensive overview of MAS theory and recent advances covering both CMASs and LMASs; (2) a dual‑dimensional comparison (theoretical and application) that highlights similarities, differences and how the paradigm shift reshapes MAS research; (3) a discussion of key research challenges and promising future directions for MAS development.

Subsequent sections provide a structured review of CMASs across perception, communication, decision and control; a systematic analysis of LMASs from architectural, runtime, adaptability and application viewpoints; a comparative analysis that elucidates the logical transition from CMAS to LMAS; and finally, an outlook that summarizes current challenges and potential research opportunities.