In the rapidly evolving field of cloud computing, the emergence of generative AI agents, or more colloquially, agentic AI, heralds a potential paradigm shift in how we do AI in the cloud—even before we fully capitalize on generative AI’s true potential.
Just as cloud computing transformed the tech landscape, agentic AI has the potential to revolutionize our approach to generative AI architecture by introducing autonomy, intelligence, and efficiency.
Before we delve deeper, it’s important to understand that agentic AI is not a one-size-fits-all solution for all AI deployments. Yes, agentic AI has mind-blowing potential. In this industry, we tend to fall for the hype of the latest hot technology without adequate understanding or experience to make informed decisions. Rather than just promoting agentic AI, my goal is to let you know that agentic AI is a viable architectural option but also to be aware of its downsides.
The autonomy revolution
At the heart of agentic AI lies its autonomy and ability to facilitate dynamic, distributed behavior. AI agents can independently initiate, plan, and complete complex tasks that traditionally require significant human intervention. Cloud architects can move from manual task management to a supervisory role where the AI handles the intricacies.
Imagine a scenario where generative AI agents autonomously manage infrastructure provisioning, scaling resources dynamically based on workload demands and optimizing configurations for enhanced performance.
The differences between agentic AI and AI agents
The term agentic AI encompasses the broader and more advanced conceptual framework. It is the overarching system with comprehensive autonomous and adaptive capabilities. AI agents are the building blocks that perform specific tasks or functions as part of the agentic AI structure. They are the operative components that execute specific tasks within this system. Agentic AI and AI agents are related but different. Clear as mud?
Agentic AI is an artificial intelligence system designed to achieve complex goals and manage workflows with minimal human supervision. It demonstrates advanced capabilities to understand context, make decisions, adapt to changing circumstances, and autonomously complete multifaceted tasks.
One critical characteristic of agentic AI is its autonomy. The AI agents (foundational to agentic AI) operate independently, initiating and executing tasks without constant human oversight. This independence allows them to efficiently carry out their responsibilities and respond promptly to various situations.
If this seems like déjà vu, you’re right. The use of agents is decades old. Once again, we’re dusting off old architectural patterns to build and define new and unique value. (Can you say “containers”?) I’ve worked with agents as an architectural option for years, including intelligent agents that use AI features. What’s new here is the use of generative AI (specifically large language models, LLMs), although it doesn’t provide that much difference.
How agentic AI works
Two crucial aspects of these agents are their decision-making and reasoning capabilities. They are equipped with sophisticated algorithms that enable them to evaluate different options, balance trade-offs, and effectively respond to novel situations. They can do this with their AI capabilities, but most will consult other LLMs to get their takes on problems they want to solve. Typically, many LLMs are consulted and then checked for consistent answers.
In addition to making decisions, AI agents are highly adaptive if appropriately built. They can adjust their actions and plans dynamically based on changing conditions and real-time feedback. This adaptability ensures that they continue to operate effectively even in volatile environments, maintaining their efficiency and effectiveness.
Agentic AI deployed in supply chain management can handle various logistics operations autonomously, ensuring that goods are transported, stored, and delivered efficiently. These AI agents analyze and coordinate data from multiple sources, such as inventory levels, delivery schedules, and real-time weather conditions.
Let’s say a global retail company utilizes agentic AI to manage its supply chain operations across multiple regions. How will it handle severe weather conditions that cause unexpected disruptions across several distribution routes? Or a pandemic? In the weather instance, AI agents would quickly analyze real-time traffic updates, weather forecasts, and port closures. Then they would dynamically adjust the delivery routes, rerouting trucks to less affected areas to avoid delays and keep deliveries timely.
These agents are also proficient at pursuing complex goals. They can handle intricate, multistep processes and workflows, setting and achieving sub-goals to accomplish any number of objectives. They can manage complicated tasks that would otherwise require significant human intervention.
AI agents possess advanced natural language processing (NLP) capabilities. They can comprehend, interpret, and generate human language, facilitating easy interaction and communication with users and other systems. These agents also work alongside other AI agents or human operators in collaborative and iterative workflows. Through continuous learning and feedback, they refine their outputs and improve overall performance.
More complicated than it appears
On paper, AI agents should be in wide use today. Look at all the pros I’ve listed. The downsides are much more difficult to understand. Even though you need tools to build AI agents, the tools are all over the place regarding what they are and how to use them. Don’t let vendors tell you otherwise.
First, these are complex beasties to write and deploy. Architects who can design AI agents and developers who can effectively build AI agents are few and far between. I’ve witnessed teams announce they will use agent-based technology and then build something that falls far short of a solution for the proposed business case.
Second, you can’t put much into these AI agents or they are no longer agents. You missed the point if your AI agents are vast clusters of GPUs. The better way is to deploy AI solutions where there is not much going on within the agents. Instead, they reach out for heavier processing requirements, such as interacting with many LLMs that carry out the “real work.”
My prediction is that we’ll see many more agentic AI architectures emerge as AI and cloud architects begin to understand their value. I’ve already integrated them into several projects. My advice? Make sure everyone understands the benefits as well as the challenges. We’re learning as we go. It’s time to investigate the possibilities and start down the path of agentic AI. Good luck.
Copyright © 2024 IDG Communications, Inc.
