Multi-Agent Systems: from classical paradigms to an LFM-driven future

A new arXiv survey by Zixiang Wang and co-authors, accepted in the IEEE/CAA Journal of Automatica Sinica, systematically connects the classical literature on Multi-Agent Systems (MAS) with modern approaches that rely on Large Foundation Models (LFMs). Published on April 20, 2026, the paper is one of the first attempts to formally document a paradigm shift.

What does the classical MAS framework offer?

Traditional MAS systems organize agents along four core dimensions: perception, communication, decision-making, and control. These dimensions form a closed-loop coordination framework in which agents exchange states, negotiate roles, and execute tasks.

The classical approach relies on control-theoretic foundations — mathematically provable protocols, fixed communication schemas, and explicitly modeled states. Examples include consensus algorithms for drone swarms, auction protocols for task distribution, and Markov decision-making in stochastic environments.

The strength of this approach lies in its provability — we can mathematically guarantee convergence, stability, and optimality. Its weakness is rigidity — every new task requires new modeling.

How do LFMs change the picture?

According to the authors, LFMs enable agents to “elevate collaboration from low-level state exchange to semantic reasoning.” Instead of sending numeric vectors, agents communicate with each other in natural language, share context, and jointly reason about plans.

This brings adaptive coordination strategies — agents can improvise protocols ad hoc, negotiate roles without pre-defined schemas, and leverage the full knowledge of the foundation model during collaboration. The paper notes that this shift is analogous to the transition from assembly language to natural language in human communication.

The survey compares classical and LFM approaches across architecture, operational mechanisms, adaptability potential, and real-world application scenarios.

What are the open research directions?

The authors identify several key challenges for the future of the field. The first is verification and reliability — LFM-based agents lose the formal guarantees of classical protocols, making it difficult to prove they will not stall or produce harmful outcomes. The second is communication scalability — natural language is expensive to exchange in real time among hundreds of agents.

The third challenge is emergent behavior — coordination can produce unexpected phenomena that are not explicitly programmed. The survey treats this as both an opportunity (adaptability) and a risk (unpredictability). Practical application in systems such as autonomous fleets or distributed AI assistants will require combining both approaches — the semantic reasoning of LFMs and the formal guarantees of classical protocols.

The paper is a useful reference guide for anyone working on LLM-agent frameworks such as AutoGen, CrewAI, or LangGraph, as it places their ideas in the context of decades of MAS research that preceded them.