AI agents are a growing priority for enterprises, with many companies interested in deploying them for a wide range of purposes, from software development and marketing to sales and customer support. Most discussions revolve around single AI agents. However, Gartner has seen a 1,445% surge in inquiries about multi-agent systems (MAS) from Q1 2024 to Q2 2025.

Despite this growing interest, many teams lack a clear understanding of what a multi-agent system entails. Below, we’ll explore what an MAS is and its core components. We’ll also highlight real-world MAS deployments across several industries.

Defining Multi-Agent Systems

A multi-agent system consists of multiple autonomous AI agents that work together to complete tasks or solve problems as defined by a user. Each agent has its own specialty and defined role, allowing it to perform actions that contribute to a common goal.

A typical MAS operates through orchestration, in which complex tasks are broken down into structured agentic workflows.

Components of a Multi-Agent System

Multi-agent systems typically have the same core components. These span the types of agents within the system, the underlying architecture, memory, data, and other areas.

Worker Agents

Worker agents are agents that act as individual contributors. Each agent is built to perform a narrow task autonomously but within defined boundaries.

For example, a grocery inventory management system might have various worker agents for different purposes:

• Agent #1: Reads and monitors stock levels for every product in the store.
• Agent #2: Scans perishable items like produce and dairy products to generate a spoilage risk report.
• Agent #3: Looks up current delivery lead times and minimum order quantities for suppliers of any item flagged as low stock by Agent #1.
• Agent #4: Analyzes point-of-sale transaction data to calculate how quickly each product sells.
• Agent #5: Monitors incoming deliveries and checks received quantities against the original purchase order.

Worker agents are often loosely coupled by design for high scalability and adaptability. They may also delegate tasks to other agents in the system when needed.

Orchestrator Agents

Orchestrator agents, also called leader agents, manage and coordinate worker agents. Leader agents can be deterministic, following hard-coded steps. They can also be dynamic, making decisions in real time to direct when agents are activated for each task.

For instance, in the above example, there may be an Inventory Orchestrator Agent that coordinates the worker agents and makes decisions about grocery inventory. Alternatively, in a software engineering context, a lead agent might coordinate the handoffs to various subagents for tasks like scaffolding, testing, or bug fixing.

Execution Environment

An execution environment is a shared space where AI agents do all their work. The environment is the place where agents interact with each other, obtain resources, and have restraints. This space can be virtual, like a computer simulation, or physical, like factory robots.

Shared Memory and Context

Multi-agent systems can’t operate successfully without the agents sharing memory and context. Shared agentic memory is typically done through a JSON object or session state. Context can be exchanged through a variety of mechanisms, such as message passing, thread passing, and common data stores.

These systems need shared memory and context so that agents can see and understand what other agents have done and then build upon that work.

Note: These shared components should not be confused with an LLM’s context window, which essentially serves as the AI model’s working memory.

Protocols and Tools

AI agents rely on protocols and tools to communicate, collaborate, and complete tasks. Many of the tools that agents use are driven by APIs, which they use to access various services and data sources.

Emerging protocols like Model Context Protocol (MCP) and Agent2Agent Protocol (A2A) are becoming more common in systems powered by AI agents. MCP enables agents and LLMs to connect to tools and data, while A2A facilitates agent-to-agent collaboration.

Policies

Multi-agent systems typically have various policies in place for a variety of areas to ensure each agent operates within specified guidelines. These areas include interaction and communication, access and security, coordination, decision making, safety, and compliance.

AI agents are increasingly being deployed not as isolated tools but as networked actors that share context, call tools, pass data, and delegate tasks to other agents. They form the foundation for agentic systems that perform work autonomously across an entire organization.

Examples of Multi-Agent Systems

Companies in a wide variety of industries are developing and deploying multi-agent systems. Here are a few real-world examples.

Infrastructure

NTT Data recently launched its Software Defined Infrastructure (SDI) Services Agent. The company describes this MAS as “conversational agentic service experience for enterprise infrastructure.” It acts as an orchestrator, deciding when to trigger other agents to complete tasks in the background. The company built the system to provide enterprises with actionable intelligence on complex, multi-vendor hardware and software infrastructure environments.

Pharmaceuticals

Madrigal Pharmaceuticals built a multi-agent platform for Pharma research and intelligence using LangChain and LangSmith. Each agent performs a specific, narrow function, such as search, analyze, and synthesize. When the orchestrator agent receives a task, it decides what should happen next, including which capabilities are required and which agents should run. The system automates and accelerates processes for creating solutions that treat MASH patients.

Healthcare

Fujitsu is developing an AI agent platform for Japan’s healthcare sector. This healthcare MAS has an orchestrator AI agent that coordinates multiple specialized healthcare-specific agents. Each agent performs a specific task, such as data structuring and interoperability monitoring. This agentic platform is designed to help improve the efficiency and stability of medical service provision.

The Case for Multi-Agent Systems

Enterprise teams are increasingly turning toward multi-agent systems because they allow them to automate complex workflows at scale. Consider the real-world examples we’ve highlighted.

NTT Data built an MAS that fundamentally transforms the way enterprises interact and manage their infrastructure. Madrigal Pharmaceuticals’ agentic system transforms data into meaningful results for MASH patients. Finally, Fujitsu aims to use systems with multiple agents to rapidly spread world-leading medical operational practices to improve Japan’s healthcare industry.

As more teams find innovative ways to deploy them, multi-agent systems are poised to become a cornerstone of enterprise transformation.