From AIOps to AgentOps: Why Agentic AI Is Forcing a Rethink of Cloud Architecture

Agentic AI Introduces a New Execution Model

Over the past year, agentic AI systems have moved well beyond research prototypes. They are now appearing in serious pilots and early production deployments across software engineering, operations, and business automation. Unlike traditional applications, these systems do not simply respond to a request and return a result. They persist, reason about goals, call tools, coordinate with other agents, and continue working long after a user closes a browser tab. In many environments, they are starting to behave less like applications and more like autonomous workers inside the infrastructure.

The cloud platforms we rely on today, however, were not designed for this kind of behavior. They were built for short-lived, stateless services that follow predictable request-response patterns, assumptions that remain central to modern cloud-native design. IEEE’s overview of cloud-native architecture principles captures these foundations well.

As more teams experiment with agentic systems at scale, those assumptions are beginning to strain. The result is not a minor tooling problem, but a deeper shift in how software behaves in production.

Why Agentic AI Challenges Traditional Cloud Assumptions

Modern cloud architecture is optimized for bounded execution. Requests arrive, work is performed, and resources are released. Autoscaling decisions rely on stable traffic patterns and service-level metrics, a model that has served microservices and distributed systems well for more than a decade. IEEE’s analysis of DevOps and cloud-native architecture reflects this operating model clearly.

Agentic systems operate on a very different rhythm. An agent may initiate work without any inbound request, maintain long-running execution loops, spawn additional agents, and invoke external tools continuously. Its workload is driven by evolving goals rather than external traffic. Resource consumption expands and contracts as its reasoning process unfolds.

This breaks several core assumptions at once. Execution is no longer neatly bounded by request lifetimes. Resource usage becomes difficult to forecast. Behavior is no longer strictly deterministic. Traditional cloud primitives remain essential, but on their own they no longer provide sufficient control.

Emerging Failure Modes Traditional Monitoring Misses

As agentic systems move closer to real production use, new classes of failure begin to surface. One example is the runaway execution loop: an agent repeatedly refines a plan and calls tools without making meaningful progress. These loops may never trigger conventional error conditions, yet they quietly consume compute, time, and budget.

Another pattern is the tool-call storm. An agent may discover an effective solution by parallelizing work across dozens or hundreds of tools or sub-agents. While logically valid, the resulting burst of activity can overwhelm downstream systems and destabilize environments.

Perhaps the most difficult problem to observe is intent drift. The infrastructure appears healthy by every metric, yet the agent’s behavior slowly diverges from the original objective. This challenge is increasingly discussed in governance research on advanced AI systems, including the Stanford AI Index’s section on emerging policy and risk challenges.

The Case For A New Layer: AgentOps

These issues point to a missing layer in modern operations. DevOps, MLOps, and SRE remain essential, but they were not designed for persistent, autonomous, goal-driven systems. One useful way to frame what is missing is through a new operational discipline: AgentOps.

AgentOps focuses on managing the lifecycle, behavior, and impact of agents in production. It introduces explicit control over agent goals, execution boundaries, resource budgets, and governance policies. Instead of treating agents as ordinary services, it treats them as semi-autonomous actors operating within carefully defined constraints.

This framing aligns with emerging industry thinking on agentic systems, including the World Economic Forum’s technical analysis of AI agents and their governance requirements.

What An AgentOps Stack Looks Like

This layer introduces an agent scheduler, an intent registry, a policy and budget engine, and agent-aware observability that traces behavior across goals and tools. Human-in-the-loop controls allow operators to intervene when the system encounters uncertainty or risk.

How Engineering Teams Should Prepare

Engineering organizations working with agentic systems should begin preparing now. Autoscaling strategies must evolve beyond traffic metrics to include agent budgets and execution depth. Observability pipelines must capture agent intent and tool usage, not just service latency and error rates, consistent with the guidance of the NIST AI Risk Management Framework.

Teams should also rethink reliability. In agentic systems, reliability includes behavioral correctness and goal alignment, not just uptime. Incident response must support suspending or overriding agent decisions. Governance models must treat agent behavior as production-critical code.

At the policy level, international bodies such as the OECD AI Policy Observatory and the OECD AI Principles on trustworthy AI emphasize these governance challenges as autonomous systems grow in complexity and societal impact

From Cloud-Native To Agent-Native

Cloud computing transformed software by making infrastructure programmable, elastic, and global. Agentic AI is poised to trigger a similar transformation in how systems reason, act, and evolve. But this next phase cannot be built on yesterday’s operational assumptions.

The cloud was built for services. The next decade will be built for agents.

The teams that invest early in AgentOps may well help define how safely and successfully that future unfolds.

About the Author

Nitin Ware is a lead engineer at Salesforce and a Senior Member of the IEEE, specializing in ML platform infrastructure, Kubernetes-based model serving, and large-scale reliability engineering. His work spans service mesh architectures, cloud-native observability systems, and sustainable DevOps practices. He writes about distributed systems, AI-driven infrastructure, and the future of scalable cloud-native platforms.

Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE’s position nor that of the Computer Society nor its Leadership.