Datumo’s R&D team recently pointed out that most LLM applications are shifting from the traditional RAG (Retrieval-Augmented Generation) approach toward a MAS (Multi-Agent System) model. They also predict that MAS evaluation will soon become the new standard for assessing AI reliability.
Today, let’s dive into what MAS really is—and how it differs from RAG.
A Multi-Agent System (MAS) is a decentralized computing framework where multiple autonomous agents interact within a shared environment to solve complex problems or achieve specific goals through collaboration, competition, or coordination.
Each agent operates independently, making decisions based on its own goals, knowledge, and perception of the environment.
Because MAS distributes problem-solving across multiple agents—rather than relying on a single agent—it offers greater scalability, flexibility, and robustness, making it ideal for tackling challenges that are too complex for a single model to handle alone.
Thanks to these characteristics, MAS is highly suitable for designing complex and dynamic systems.
In a single-agent system, one agent interacts with the environment, calling tools and completing tasks as needed.
In contrast, in a Multi-Agent System (MAS), multiple agents coexist and must model and consider each other’s goals, memories, and plans.
This setup makes both direct and indirect communication between agents essential. Depending on the situation, agents either cooperate or compete to solve problems.
Under a Multi-Agent System (MAS), other agents are part of the environment itself, making learning and coordination significantly more complex.
When agents learn autonomously, they typically rely on Multi-Agent Reinforcement Learning (MARL). In MARL, multiple agents simultaneously perform reinforcement learning within a shared environment, each developing optimal policies through interaction and feedback.
For instance, in a drone swarm where each drone learns its own search strategy, the movements and actions of other drones must be treated as dynamic environmental factors. Each agent independently engages in policy learning, receiving rewards and continually refining its behavior.
Critically, because other agents are constantly influencing the environment, the environment remains non-stationary—it does not stay fixed.
Key challenges in MARL:
Non-Stationarity:
The learning target keeps shifting because the environment, driven by other agents, is constantly changing.Credit Assignment Problem:
When a team succeeds, it’s difficult to individually attribute credit to each agent’s contribution.Partial Observability:
Agents have limited visibility and must act based on incomplete information about the environment.
Q. Does MAS require multiple physical robots?
A. No. MAS can also be built with software agents (programs), and most early-stage development typically takes place in simulated environments.
Q. Without centralized control, isn’t it hard to manage errors in MAS?
A. Yes, it is. That’s why MAS designs must include robust error detection and recovery mechanisms among agents.
Q. What skills are needed to apply MAS?
A. You’ll need knowledge of agent-based modeling, an understanding of network communication protocols, a foundation in reinforcement learning (especially MARL), and expertise in security design.
Q. What tools are commonly used to quickly experiment with MAS?
A. Popular frameworks include Mesa (Python-based) and JADE (Java-based).
