This article has been written and paid for by Microsoft. The content is not WIRED editorial content. The content does not necessarily reflect the views of WIRED, its affiliates, or owner, and does not reflect any endorsement direct or indirect of Microsoft, its affiliates, or other clients.
The next frontier of AI is coming into focus as agents take on longer, more strategic work. They can write and validate code, coordinate across systems, rebalance inventory as conditions change, manage production operations, or carry a migration from plan to completion. AI is no longer supporting one step; it is executing and reshaping the process itself.
New models help drive this change by expanding what AI can reason about, create, and accomplish. But the combination of advances in models and infrastructure is key to moving agents from isolated tasks to sustained, high-impact work. Realizing that potential requires silicon, compute, storage, and networking to work together so workloads run faster and more reliably at scale.
Agents Are a Different Kind of Workload
Cloud infrastructure has evolved before to support new kinds of applications. Even the most advanced cloud-native applications often followed predictable request-and-response paths. To deliver this at scale, cloud-native architectures turned to containers, distributed services, and dynamic scaling, to make the platform more automated and flexible. Agentic applications build on that foundation but introduce another shift: long-running work whose path and resource needs are dynamic and often unpredictable.
An agent may work for hours or days, moving between reasoning, data retrieval, tool use, and code execution while preserving progress between steps. Its needs may shift throughout the task. It might require a burst of computing power for one step, fast access to large amounts of data for another, and then pause while another tool or another agent finishes its work. Supporting this ebb and flow requires the platform to evolve again—this time around how agents operate.
Infrastructure Must Work as One System
Every layer of infrastructure is evolving for AI. Compute must respond to the changing demand of agentic workloads. Storage must preserve progress across long-running work. And networking must move data quickly among models, tools, and business systems. But improving each layer independently is no longer enough.
Agent performance depends on how well those layers are evolving and how well they all work together. More computing power will not help if the platform cannot move data quickly or maintain progress between steps. Silicon, compute, storage, and networking need to be co-designed and continuously optimized as one system.
This integrated approach also changes infrastructure operations. Teams need a complete view of performance, capacity, and cost so they can find bottlenecks and improve the full workload instead of tuning one component at a time. That system-wide view helps turn infrastructure improvements into faster, more reliable, cost-effective workloads.
Different Workloads, One Operational Approach
That consistent operational view becomes even more important as teams choose different technologies for different workloads. A customer-facing agent may prioritize low latency, while a complex workflow may need deeper reasoning. Teams need the freedom to choose the right model, run time, silicon, and environment for each workload while balancing performance and cost.
But more choice can create operational complexity because each combination may require a different approach to deployment, monitoring, and optimization. An integrated platform preserves flexibility while giving teams a consistent way to operate across models and infrastructure. Teams can choose what works best without managing every combination differently.
Levi Strauss & Co. And the Value of an Integrated Platform
Levi Strauss & Co. began its path to AI transformation by preparing existing business systems for the future. The company moved its data centers to Microsoft Azure, consolidated ERP platforms, modernized applications, and unified its data in a governed, real-time layer. But that was only the beginning.
Levi Strauss & Co. is now building a unified “super agent” to coordinate more than 1,000 domain-specific agents. Its platform must support both the critical processes that run the business today and a new class of intelligent workloads that coordinates work across functions. The lesson is clear: A durable platform must be built for current applications and whatever AI makes possible next.
Three practical tests for an agentic platform
Technical leaders can begin with three questions to assess readiness for an agentic platform.
First, do I have the high-performance compute, storage, and networking that models and agents need—or am I trying to run tomorrow’s applications on yesterday’s platform?
Second, can I choose the right model, run time, silicon, and location for each agentic workload while balancing performance, cost, complexity, and risk?
Third, can I use agents to operate the platform by moving workloads into production faster, resolving incidents before disruption, and continuously improving performance, reliability, and AI economics?
The next decade of AI will not be defined by a single breakthrough. Models will become even more capable, agents will take on work we have not yet imagined, and infrastructure demands will shift with them. Organizations need a durable platform that can embrace each new wave of innovation without being rebuilt whenever the frontier moves.
Azure is the platform foundation for that future. Microsoft has spent years bringing some of the world’s most advanced and demanding AI solutions online at massive scale. The lessons from that work continually shape Azure silicon, compute, storage, networking, and operations—and make those capabilities available to customers.
This is the opportunity of Frontier Transformation: continually turning advances across the AI stack into new ways to operate, compete, and grow.
Learn more at azure.com.
The post The Platform Behind the Next AI Breakthrough appeared first on Microsoft Azure Blog.
Quelle: Azure
Published by