Although generative AI still feels new and innovative, we’re already approaching the two-year anniversary of the launch of ChatGPT. It’s safe to say that we’re now firmly in the GenAI era. In fact, it’s been long enough that some people may be starting to wonder whether GenAI has lived up to the considerable hype, and if not, why not?
There’s no question that GenAI is capable of great things, but it isn’t enough to enable many enterprise-class use cases. Current GenAI workflows are limited in several different ways. That’s why a new class of AI, known as agentic AI, is emerging. Agentic AI is complementary to GenAI, and helps take GenAI workflows to the next level in the evolution of AI.
The field of intelligent agents has been around for decades, but now, the synergy of advanced GenAI algorithms with agent technology can solve real business problems for enterprises. Similar to how many providers (including clouds and Model as a Service providers) are now offering AI models that others can use for fine-tuning or retrieval-augmented generation (RAG), we’re entering the realm of agentic AI blueprints (also known as templates or workflows) that automate common tasks like flight booking, customer service and virtual screening for drug discovery. These blueprints can be industry-agnostic or industry-specific and are offered by clouds and service providers.
The following table captures how the combination of GenAI and agents brings value to enterprises.
| Required capabilities | Current GenAI architectures | GenAI + agentic AI |
| Access to dynamic/real-time data | LLMs are typically trained on static datasets that include information up to a certain point in time. RAG architecture is intended to bring more dynamism but still isn’t sufficient. | Agents can dynamically pull real-time data from multiple external providers such as stock market, weather and traffic data to complement GenAI query results. |
| Goal-oriented | Most current GenAI models target content generation use cases. | Agentic AI architectures are more goal-oriented: Multiple agents deploying multiple AI models can collaborate to complete a specific objective. |
| User interface flexibility | LLMs provide greater flexibility with respect to natural language interfaces than traditional programming, but can’t support critical processes that require precision and predictability. | Agents provide both the flexible interfaces of GenAI and the precision of structured queries and APIs. |
| Autonomous operations | GenAI models typically have a human in the loop to help guide the decision-making process. Therefore, they tend to operate at human latency levels. | Agents actively seek new data and refine themselves to adapt to environmental changes, with or without a human in the loop. In many instances, these agents interact with each other at machine latency levels. |
Agentic AI is the next major step in our AI journey. It can help enterprises enable use cases that have been out of reach until now. This fact will likely drive rapid adoption. According to Gartner®:
“By 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024, enabling 15% of day-to-day work decisions to be made autonomously.”[1]
The key takeaway from the above table is that agentic AI workflows will consume data and AI models dynamically from multiple external providers.
What makes a workflow agentic?
An AI workflow is agentic when it has the ability to act as an intelligent agent. This means that it can perceive the environment around it and take specific actions based on what it perceives. The fact that it has agency signifies that it doesn’t need to be prompted by a human in order to do this.
Agentic AI is a new development based on an established concept. Back in the 1990s, IBM introduced the MAPE control loop—monitoring, analysis, planning and execution—as a blueprint for autonomic computing.[2] The feedback loop includes systems that constantly monitor datastores and retrain models as needed. In some cases, humans still play a role in fine-tuning the loop and giving the final say on the overall execution. This is known as human in the loop (HITL).
Figure 1: Agent architecture (MAPE loop)
In the machine learning era, the analysis and planning components of the MAPE loop have grown far more sophisticated, incorporating both GenAI and conventional models. This is why agentic AI workflows can support use cases that require both flexibility and precision. At the execution phase, agentic AI workflows can trigger themselves without the need for HITL. This includes working with a hierarchy of other intelligent agents to accomplish common objectives.
Agentic AI workflows can apply different tools and resources to meet these objectives. This includes pulling real-time data directly from many different sources, including sensors, databases or APIs. After finding relevant insights, they’re able to reason about which action to take next. They can then learn from the actions they take, compiling data to help improve their decision-making in the future.
Figure 2: Agentic AI pipeline design patterns
Each of the dots in the figure above represents an agent. An AI model can be encapsulated inside an agent. The agents in a pipeline can come from the same provider or different providers, and they can execute in the same location or remotely at a different location. Thus, the placement of these agents (where they execute and their location relative to data sources and AI models) and the network connectivity between them matters.
Agentic AI use cases
The potential applications for agentic AI in enterprise settings are practically limitless. We’ll look at just a few examples here.
High-frequency trading
With stock market data now widely available in real time, the race to capitalize on favorable trades happens quicker than the blink of an eye. Thus, acting without delay—whether that delay comes from network latency or human intervention—is essential.
High-frequency trading (HFT) firms can design their intelligent agents to seek out market inefficiencies and then execute trades before the opportunity disappears in nanoseconds.
Healthcare
Intelligent agents are able to constantly monitor sensor data to detect even seemingly minor changes in a patient’s condition. When they detect such a change, they’re able to reason which adjustments should be made to the care plan. Crucially, they can decide which changes are significant enough to require the expertise of a human doctor. In cases like this, they can alert the doctor immediately.
Marketing
Agents can autonomously monitor the effectiveness of marketing campaigns by tracking both external social media networks and internal sales results. They can then take corrective actions to improve performance.
Human resources
Agents can dynamically monitor the qualifications and capabilities of employees and new hires via emails, chats, support tickets and presentations. They can then adjust the required employee training by assigning courses or connecting employees with SMEs and mentors.
Supply chain management
Companies are leveraging agents to automate inventory monitoring. By automatically placing orders, predicting supply chain disruptions based on external events and adjusting production schedules, they can maintain optimum inventory levels.
Cybersecurity
Agents can continuously monitor employee behavior, network traffic, emails and system logs to detect security threats. They can then dynamically take corrective actions such as sending alerts or controlling access.
Agentic AI requires infrastructure in the right locations
Like any other form of AI, agentic AI requires distributed digital infrastructure that’s designed and strategically positioned to meet the needs of different workflows. In fact, since agentic AI promises more advanced capabilities than traditional AI, it’s no surprise that it requires more advanced infrastructure to deliver on that promise. Figure 3 shows how enterprises can deploy agents on a distributed infrastructure platform to leverage the business value described below.
Figure 3: Equinix value proposition for agentic AI
Where you host your agentic AI pipeline assets—both the agents themselves and the datasets those agents consume—is important for several reasons.
Deploying in AI-ready data centers
In many use cases, agents will need to execute on AI infrastructure that consumes a lot of power and requires special cooling (for example, service provider solutions that instantiate millions of agents on behalf of their customers). As shown in Point 1 in Figure 3 above, Equinix data centers allow enterprises to deploy next-generation AI racks that can consume 40KW+ per rack and require liquid cooling.
Keeping latency low
As mentioned above, agentic AI is ideal for use cases that operate at machine speed, including high-frequency trading. The value of these use cases comes from operating at speeds much faster than a human being could even begin to register, but agents can’t do that if they’re delayed by latency. Thus, as shown in Point 5 in Figure 3, it’s essential for enterprises to deploy agents close to where the data is generated or the end users are located. It’s also important that those agents be connected via high-speed, predictable private interconnection.
Ensuring cost-efficiency
Despite the clear potential of agentic AI, it may be difficult to secure executive buy-in for use cases that aren’t cost-efficient. Thus, as shown in Points 3 and 4 in Figure 3, enterprises need distributed digital infrastructure that allows them to process data in the metro where it was generated. This helps avoid data backhauls that consume a lot of network resources and therefore increase costs. Also, they must be careful to reduce cloud data egress fees as they design their hybrid cloud agent architectures.
Meeting privacy and compliance requirements
Agentic AI can’t be a data free-for-all. Enterprises need to ensure their agents can pull the data they need without putting that data at risk of exposure. In addition, they need to know they’re not using data somewhere they shouldn’t be. For instance, they may have to comply with data residency regulations that require certain types of data to remain within a particular jurisdiction. Therefore, as shown in Point 4 in Figure 3, they need their agents to be hosted in that same jurisdiction, so that the data doesn’t have to move in order to reach the agents.
Maintaining choice and flexibility
Agents need to pull data from different sources and communicate with other agents hosted on different provider platforms. Therefore, as shown in Points 1 and 3 in Figure 3, it’s beneficial to have a copy of your storage or data lake (authoritative data core) and your coordinator agents deployed at a neutral infrastructure location like Equinix. This ensures the flexibility needed to access agents from multiple providers and protects against vendor lock-in. Thus, it allows enterprises to pick the right agent for a particular task without worrying about which platform the agent is hosted on and to leverage best-of-breed agents from different providers.
The right partner for agentic AI infrastructure
By meeting each of the infrastructure requirements outlined above, as shown in Figure 3, Equinix is uniquely positioned to enable agentic AI adoption.
We offer a global platform of data centers and interconnection services to help our customers deploy at the edge, keeping latency and network costs low.
Our platform is vendor-neutral by design, and we offer low-latency cloud on-ramps to all major cloud providers. Our customers can deploy adjacent to the cloud regions of their choice. This means they can access cloud services on demand while also avoiding vendor lock-in and data egress fees.
Finally, we can help enterprises pursue a private AI strategy, where they build and run models on private infrastructure only. This allows them to strategically position agents and datasets in the right locations to ensure privacy and compliance with data residency regulations.
Learn more about what Private AI with Platform Equinix® can help you accomplish: Read our joint e-book with NVIDIA.
[1] Gartner, Intelligent Agents in AI Really Can Work Alone. Here’s How. Tom Coshow, October 1, 2024.
GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.
[2] Steve R. White, James E. Hanson, Ian Whalley, David M. Chess, Jeffrey O. Kephart, An architectural approach to autonomic computing, IBM Research, September 27, 2004.
Subscribe to the Equinix Blog