Autonomous AI agents are fundamentally changing the landscape of enterprise computing, a development explained by Deanna Berger on IBM Technology. As Deanna Berger noted, users are quickly realizing that the strategic integration of AI's diverse capabilities into "holistic solutions" is the key to unleashing the full power of the technology. The traditional difficulty of mapping a team's goals into a seamless, end-to-end AI strategy, where components "work seamlessly together," can be "puzzling," but this challenge is being resolved by the introduction of AI agents.
Deanna Berger detailed why these AI agents are so powerful. Unlike traditional AI models that are reactive and predictive, AI agents are initiative, goal-driven, and context-aware. A crucial feature of these systems is their ability to maintain short- and long-term memories, which they use to learn, reflect, and adjust future behavior. Critically, they can plan complicated, multi-step workflows. Once a plan is formulated, they can interact within the "metaverse," or software ecosystem, to execute that work autonomously, meaning they can "make a plan and then they can get things done".
This autonomous capability is essential for overcoming the complexity inherent in AI infrastructure transformation. Agentic AI can "autonomously assemble" the puzzle pieces required to form more useful solutions, eliminating the manual challenge of combining use cases with the optimal mix of different models, connecting necessary software, and integrating the growing use of AI accelerator cards. The resulting benefits include "much, much higher inference and decision accuracy" and "much lower overhead and operational cost," because so much is not being done manually by the human team. As a result, the productivity of both the team and the AI solution itself goes up.
Related article - Uphorial Shopify
]
The power of these agents is amplified by their ability to interact widely within the software ecosystem, or "metaverse". An AI agent can interact with customer applications via APIs, access databases to retrieve data and run models, and utilize "basically any resource that's available out in the cloud". Furthermore, they interact with the software and the actual computer running the agent program, and in some instances, even the firmware for specific accelerator cards. Deanna Berger's "favorite piece of this" architecture is the capability for AI agents to interact with other AI agents. These secondary agents can take on a specialized task, communicate with the ecosystem to complete it, and then return control back to the original agent.
To fully illustrate this, Deanna Berger presented a simplified, real-world example of a car insurance company modernizing its claim process. A specialized AI claim agent could be created. Because it is agentic, it would plan and execute a multi-step workflow for the claim data. This workflow might involve parsing the claim, matching it with the policyholder’s information, performing image processing, running fraud detection, completing audit work, and handling client interaction. The claim agent uses its understanding of the ecosystem to determine the most effective approach for each step.
For example, to parse the claim, the agent might outsource the task to a good NLP model available in the cloud. Policy matching, which is well-suited for a large language model (LLM), could be offloaded to a PCIe-attached acceleration card on an enterprise system. Image processing, perhaps to determine if pictures are fraudulent, could be sent to a bank of GPUs. Fraud detection might be handled by sending the model to a card with optimized firmware if available in the ecosystem. The agent would interact with a database for the required audit step.
For the final step—communicating the results back to the client—the claim agent recognizes that the necessary skills are "vastly different" from those used for processing the claim. Deanna Berger suggested this is an ideal use case to create a second AI agent—a client interaction agent—because the skills needed to process claims and interact with clients are different, much like they are for a human. This specialized agent would possess the tools and capabilities to interact effectively with client applications and databases to communicate the results back to the user. Deanna Berger concluded that while terms like "metaverse" and "agentic AI" may seem mysterious on the surface, breaking the concept down into its cooperative components is the fundamental key to understanding the real power of this technology and sparking inspiration for future possibility.
