Multi-Agent Systems: A Gentle Intro

Published: December 18, 2025

What Is an Agent?

Before discussing multi-agent systems, we must first answer a deceptively simple question: what is an agent?

In classical artificial intelligence, an agent is an entity that perceives its environment through sensors and acts upon that environment through actuators. This definition, popularized by Russell and Norvig, emphasizes interaction, autonomy, and decision-making under uncertainty.

Importantly, an agent is not defined by its internal implementation. It may be rule-based, learning-based, symbolic, or statistical. What matters is that it continuously participates in a perception–decision–action loop.

Agents in the Age of LLM(s)

In today’s AI ecosystem, the term agent is frequently applied to systems that are little more than scripted workflows around large language models. A typical example involves an LLM calling tools, executing steps in sequence, and returning a final response.

While such systems may be useful, they differ fundamentally from agents as defined in the AI literature. Most lack:

Calling these systems “agents” is often a matter of convenience or marketing, rather than a reflection of their underlying architecture.

What Is an Intelligent Agent?

An intelligent agent is not merely reactive. It selects actions that maximize expected performance relative to a goal or utility function. Intelligence, in this sense, refers to rational behavior, not human-like cognition.

Intelligent agents typically exhibit:

What Makes a System Multi-Agent?

A system is considered a multi-agent system (MAS) when it is composed of multiple intelligent agents that interact within a shared environment. Crucially, this interaction may be cooperative, competitive, or mixed.

However, not all systems with multiple components qualify as true MAS. The defining property is decentralization.

Decentralization as a First-Class Requirement

In a genuine multi-agent system, no single agent possesses global control. Each agent operates based on local information, communicates with peers, and coordinates behavior without relying on a central orchestrator.

Once a system collapses control into a single coordinating entity, it effectively becomes a monolithic or hierarchical system, even if it contains many subcomponents.

Most “Multi-Agent” Systems Are Pipelines

Many modern systems labeled as multi-agent are better described as pipelines or workflows. These systems decompose tasks into steps and execute them sequentially under centralized control.

Property True Multi-Agent System LLM Workflow / Pipeline
Control Decentralized Centralized
Decision-making Local and concurrent Sequential and scripted
Failure mode Graceful degradation Single point of failure
Adaptation Continuous and distributed Limited or manual

Core Capabilities of Multi-Agent Systems

For a system to meaningfully qualify as multi-agent, it must support several core capabilities:

Conclusion

Multi-agent systems are not a recent invention, nor are they synonymous with modern “agentic” AI tooling. They represent a deep and rigorous field focused on decentralized intelligence, coordination, and emergent behavior.

As interest in agent-based systems continues to grow, maintaining conceptual clarity is essential. Not every workflow is a multi-agent system, and not every tool-calling LLM is an agent.

References & Further Reading

  1. Russell, S. J., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
  2. Wooldridge, M. (2009). An Introduction to MultiAgent Systems (2nd ed.). John Wiley & Sons.
  3. Shoham, Y., & Leyton-Brown, K. (2009). Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations. Cambridge University Press.
  4. Wikipedia contributors. (2025). Multi-agent system. https://en.wikipedia.org/wiki/Multi-agent_system
  5. XMPRO. (2024). Beyond LLMs: Why true multi-agent systems are 90% business process intelligence. Link

[Top]