Is One Agent Enough to Rule the World?

Widely regarded as the dawn of agentic AI, 2025 marks a pivotal point in the evolution of intelligent systems. Last year, I explored the transformative potential of multi-agent collaboration, and more recently, how human-AI teamwork will redefine how we interact and work alongside AI. As these autonomous agents become mainstream, their integration with human teams promises not just efficiency, but a profound evolution in how we approach collaboration, creativity, and productivity across industries.

Experts across the AI landscape have examined various types of multi-agent architectures—such as centralized, decentralized, and hierarchical—and the distinct benefits they offer, from scalability to adaptability. For instance, in his Medium piece, Sahin Ahmed outlines the foundational types of multi-agent architectures and their respective strengths, while Nathan Lambert’s article dives deeper into the future of agent-based systems, showcasing emerging architectures capable of autonomous reasoning and dynamic collaboration.

An intentional path for implementing AI

We’ve chosen our path with intent, resulting in several patents we filed in the fall of 2024. A key to success lies in how the plan is crafted—using a variety of reasoning and reflection techniques to ensure the right agents are assigned to the right tasks. As we’ve explored these methods, one insight has become clear: errors can quickly multiply. Even if each step or task is optimized into the 90th percentile, the overall outcome can fall significantly short if the planner agent isn’t carefully constructing an effective execution strategy. I will revisit the role of the planner agent later in the blog, as it plays a central role in orchestrating successful multi-agent outcomes.

Real-world applications in networking and security

Enterprise platforms are designed to collect, process, and store vast amounts of both structured and unstructured data from a wide range of connected devices, clients, users, and applications. This includes operational data, enterprise knowledge, and content such as service contracts, subscriptions, licensing information, documentation, manuals, release notes, and more. By integrating IT infrastructure data with broader business information, these platforms enable insights for decision-making, reasoning, task automation, and enhanced customer experiences.

These goals can’t be achieved by a simple chatbot, or single agent built from off-the-shelf components available today from most of the major providers. Their capabilities are too limited in scope, which led us to expand our initial architecture into a multi-agent system leveraging a specific set of tools tailored for network and security operations.

Our system is designed to enable autonomous network operations through goal-oriented agents capable of perceiving their environment, reasoning, making decisions, and performing actions to achieve defined objectives. These agents are built to handle complex tasks efficiently and adaptively—with little to no need for human oversight. Of course, success depends not only on the technology itself, but also on the level of trust users place in it. That’s why we expect users to want control over the degree of agency, including the ability to decide when and where human-in-the-loop involvement is required, especially before executing critical changes.

In this architecture, human users collaborate with an intelligent network of AI agents, led by a planner agent responsible for creating the overall strategy and invoking the appropriate agents to achieve the defined goals. Serving as the strategic coordinator, the planner agent dynamically manages a team of specialized worker agents, each equipped with a broad range of tools and systems. These worker agents can interact with other models and enterprise tools, call REST or GraphQL APIs, and query various types of databases—including SQL, NoSQL, and graph structures. They also engage with vector stores, search engines, and leverage live data streams to inform and optimize their actions.

Approaches to AI architectures

Depending on the task at hand, there may be different approaches you choose to take for orchestrating agent communication—centralized, decentralized, and hybrid—each optimized for different operational needs. In a centralized architecture, the planner agent coordinates every communication between the worker agents, enabling simplified management and efficient resource allocation (but adding delay as tasks are executed sequentially), especially for lower-complexity tasks requiring consistency and unified control.

In contrast, a decentralized architecture allows autonomous agents to collaborate directly, enhancing scalability and fault tolerance for large-scale or distributed tasks.  However, without a central point of coordination, these systems can quickly become difficult to manage, making it challenging for agents to converge on a consistent result or defined set of outcomes. The hybrid approach combines both models, dynamically adapting the architecture based on task complexity and network conditions—centralized oversight for strategic goals, and decentralized execution for specialized sub-tasks.

Human involvement with AI

While agentic systems operate autonomously—depending on their level of agency—they still require engagement at key moments. This involvement may be initiated by the user, triggered on a schedule, or activated by an event, which itself could be the result of continuous monitoring by a model observing specific infrastructure characteristics. An example would be a troubleshooting request to an agentic system which consists of a team of specialized troubleshooting worker agents to identify and resolve network issues. For example, the planner agent could initiate workflows involving wireless, wired, and fabric-specific worker agents, each focused on diagnosing their respective domains.

Worker agents are equipped with access to both historical and real-time monitoring data, along with specific tools to perform their assigned tasks. The tasks assigned to the agents are aggregated into a unified result that identifies potential root causes, recommends remediation steps, and—after human validation and approval—can trigger autonomous corrective actions.

Introducing Extreme Platform ONE

At Extreme Networks, our AI is embedded right from the start in Extreme Platform ONE. This intelligent system enables organizations to simplify and automate operations, enhance security, and improve user experiences by automating complex workflows across networking, security, and IT domains.

In the coming months, I’ll be sharing deeper insights into the technical aspects of this architecture, along with more practical examples that highlight how our intelligent agents drive meaningful impact in real-world scenarios.