Tuesday, 27 May 2025
Multi-Agent AI Systems
This article is part of our trend series on Multi-Agent AI. Explore all related content.
Beyond the single agent: collaborative intelligence
AI agents are the building blocks of multi-agent systems. As we’ve seen, these agents are programmed with either general or specialized functions to handle tasks of varying complexity, such as information retrieval, source validation, data formatting, and report generation. Aside from these distinctions, each agent is a system capable of interacting with its environment. It collects data and is programmed or configured to make decisions and achieve predefined goals.
Multi-agent systems have been around for over a decade, powering everything from drone swarms used for large-scale detection or planting tasks, to the thousands of robots in massive warehouses that optimize package creation.
What’s driving renewed interest in these systems is the integration of large language models (LLMs) into their design, enabling more natural interaction and adaptability.
The true potential of intelligent agents emerges when multiple agents collaborate. This create a multi-agent AI system that can be compared to the way beehives or ant colonies operate, where each member has a specific role to play (foraging, building, organizing, defending, etc.). The terms “collective intelligence” and “distributed intelligence” are used to describe how these systems function, each agent being either interdependent or autonomous depending on the task at hand. The agents can be thought of as teams of virtual experts working together toward a common goal.
Agent orchestration: The key to an effective collaboration
Orchestration refers to the coordination and management of agent actions within a multi-agent system.
There are several orchestration models, each suited to different types of problems:
- Centralized Orchestration: A single “coordinator agent” delegates tasks, monitors execution, and integrates results. This is the most commonly used model. The orchestrator allows for precise control, but it can become overwhelmed when the number of agents grows too large.
- Decentralized Orchestration: Agents negotiate directly with each other to divide work and coordinate actions. This approach is more robust, especially in the event of failures, but it’s also more complex to implement.
- Hybrid Orchestration: As its name suggests, this model combines centralized and decentralized orchestration depending on the task and its importance.
Orchestration is all the more effective when it can manage conflicts between agents and ensure the smooth execution of sequences throughout the entire process.
Real-World Applications
Finance
JPMorgan uses a multi-agent system (DeepX) where each agent analyzes different market indicators (macroeconomics, sector trends, company data) and then combines them to deliver more comprehensive and nuanced investment recommendations.
Logistics
DHL has deployed a MAS for route optimization, where each truck is modeled as an agent. These agents negotiate with each other so that a central orchestrator can optimize routes in real time, resulting in a 15% reduction in fuel costs.
Retail
Ocado, a pioneer in online grocery retail, orchestrates over 2,000 robot agents in its warehouses. Each robot moves along a grid, picking up and dropping off products. By coordinating their movements, the robots increase order fulfillment efficiency by 50% compared to traditional methods.
The Agentic Mesh
Beyond simple groups of agents, we’re now seeing the rise of the Agentic Mesh, an interconnected ecosystem that facilitates collaboration, interaction, and transactions between autonomous agents. This structure allows agents to discover each other and form dynamic connections for cooperation.
In this ecosystem, agents form temporary coalitions to solve specific problems and reconfigure themselves to tackle new challenges.
The Agentic Mesh typically includes:
- A registry that centralizes metadata about available agents
- A marketplace for discovering and engaging agents
- Standardized protocols that enable efficient communication
- Reputation mechanisms to assess agent reliability based on past performance
This approach paves the way for applications where multiple intelligent components divide the work to achieve goals more efficiently than a monolithic solution.
Multi-Agent Systems and Technological Innovation
The rise of MAS is part of a broader shift in computing toward distributed and adaptive architectures. Several factors are accelerating their adoption:
- The availability of massive datasets and powerful AI models, especially LLMs capable of understanding natural language instructions, makes it possible to equip each agent with highly specialized intelligence.
- The widespread use of distributed architectures (cloud computing, microservices, IoT) makes it easier to deploy multiple agents across enterprise systems.
- Major tech providers now offer robust frameworks for building and orchestrating these agents: AWS with Bedrock, Microsoft with Semantic Kernel, Salesforce with Agentforce, and many others.
Multi-agent systems offer a new way to tackle complex problems. By distributing decision-making across multiple specialized and coordinated entities, they offer a model that more closely mirrors how human organizations operate.
Transform your business with AI
