The Kubernetes moment for AI: the rise of the agent control plane

The industry is moving past the novelty of individual AI agents. The challenge is no longer just building a capable agent, but managing a fleet of them. We are seeing a shift from agent development to agent orchestration and governance.

In many ways, this is the “Kubernetes moment” for AI.

From logic to lifecycle

Early agentic work focused on the “how”—the logic, the prompting, and the reasoning loops required to complete a task. But as these agents move from experimental sandboxes into production environments, they bring a new set of requirements that traditional compute platforms aren’t built to handle.

Agents are different from stateless microservices. They are long-running, stateful, and often bursty. They require integrated network policies, secure credential injection, and sandboxed execution environments. If you try to manage a fleet of autonomous agents using standard container orchestration alone, you will likely end up building a custom management layer just to keep them secure and observable.

The emergence of the control plane

Dedicated infrastructure is beginning to emerge to solve this. Two recent developments illustrate the trend:

WSO2 announced the General Availability of WSO2 Agent Manager in September 2026. This open-source platform centralizes identity, permissions, and lifecycle oversight, allowing organizations to manage agents across different models and frameworks from a single point of control.

Google’s AX Open Runtime is a declarative control plane built on the Agent Substrate. AX uses specific primitives—Task, Workspace, Gateway, and Model—to manage stateful workloads. It handles the unique lifecycle of an agent, such as checkpointing and resuming operations in under a second.

These are tools for managing agents, not just building them. They represent the rise of the “Agent Control Plane.”

Why framework-agnosticism matters

For an enterprise, the goal is to avoid vendor lock-in. If your security and identity architecture is baked directly into a specific agent framework or an LLM provider’s ecosystem, you are stuck.

A dedicated control plane provides a layer of abstraction. By separating agent logic from governance, organizations can swap out LLMs (moving from OpenAI to Anthropic or a local model) or change frameworks (moving from LangChain to CrewAI) without rebuilding their entire security and identity posture. The control plane handles identity, permissions, and auditing, while the agent focuses on the task.

The complexity trade-off

This shift toward specialized infrastructure has friction.

For small-scale deployments, like a single agent running a single workflow, a full-scale control plane is likely unnecessary architectural overhead. There is a risk of over-engineering for problems that do not exist yet.

There is also tension between fragmentation and consolidation. While independent control planes offer the freedom to mix and match, major cloud providers like AWS and Microsoft are already integrating agentic governance directly into their platforms. This could eventually make independent control planes harder to maintain.

Finally, this infrastructure is still in its infancy. The primitives and identity protocols we are seeing today are still evolving.

Bounding the blast radius

As agents gain more agency—the ability to act on real systems, move money, or access sensitive data—the stakes of management increase. In this context, the control plane is a critical defense-in-depth layer.

If the runtime handles the execution of the agent, the control plane defines its boundaries. It is the primary mechanism for bounding the blast radius of autonomous systems, ensuring that as agents become more capable, they also become more controllable.