Container Virtualization DevOps

Kubernetes Projects Target Agent Workload Efficiency and Security

Kubernetes is facing new scaling and security demands as enterprises begin running long-lived artificial intelligence agents on clusters. At KubeCon + CloudNativeCon North America 2026, discussions highlighted the resource costs of keeping agents available between requests, the resulting load on the Kubernetes control plane, and the need for stronger safeguards around agent activity.

Several initiatives are under way, including Google’s Agent Substrate framework, Google’s gVisor isolation technology and NVIDIA’s OpenShell runtime. Their contributions address different challenges, but the work does not mean that current Kubernetes deployments are already optimized for agentic AI. Integrating these approaches into the environments used by IT teams will take time, and the broader platform may need further changes.

Idle agents create a resource and control-plane burden

Many agent workloads can remain inactive while waiting for a model response, an external service or human input. In the current deployment pattern described at the conference, each active agent receives its own Kubernetes pod. Keeping that pod allocated during a pause can tie up CPU and memory that could otherwise serve other workloads.

The cost is not limited to the workers running those pods. Kubernetes must also track the persistent objects associated with them. At very large scale, maintaining millions of objects can add work for a control plane that was not designed around vast numbers of long-lived, intermittently active agent sessions.

Agent activity may also increase the number of calls made to Kubernetes APIs. Mickey Boxell, senior principal product manager for cloud native services at Oracle, pointed to control-plane performance as a concern as AI agents generate more API requests. The challenge, therefore, spans both efficient use of compute resources and the capacity of Kubernetes management components to handle the resulting activity.

Agent Substrate separates session state from compute

Google has contributed Agent Substrate, an open source framework, to the Cloud Native Computing Foundation. The framework is intended to let long-running agent sessions retain their state and identity without requiring their compute resources to remain continuously allocated.

When an agent becomes idle, Agent Substrate can save a snapshot of its state and release the worker for other tasks. When the session is needed again, the framework restores it on available compute. Google says restoration can take less than 500 milliseconds. That figure is the company’s stated capability; the source material does not specify the conditions or workload under which it was measured.

The approach aims to reduce the infrastructure consumed by agents that spend substantial time waiting. It also offers a different way to manage session continuity than keeping a dedicated pod active throughout each pause. The framework’s contribution to the CNCF makes it part of the open source cloud-native ecosystem, but does not by itself establish when or how broadly it will be incorporated into Kubernetes distributions.

Security efforts address isolation and agent permissions

Resource efficiency is only part of the problem. Agents can execute code and interact with services, creating a need to limit both the risks of untrusted workloads and the actions an agent is allowed to take.

Google’s gVisor provides isolation by running code inside a lightweight userspace application kernel that can be deployed as a container. Its role is to help separate untrusted code from the host environment. NVIDIA has also contributed OpenShell, an open source runtime for securely deploying AI agents, to the CNCF. OpenShell focuses on enforcing policies governing what agents can access and do. These technologies address related but distinct security concerns: workload isolation and controls on agent behavior.

At the conference, Alex Zenla, CTO of Edera, said Kubernetes has demonstrated that it can be extended to address architectural needs. He pointed to the platform’s evolution to support stateful applications, despite its origins in managing stateless workloads. That history suggests a possible path for adaptation, but it does not remove the engineering work required for agent workloads.

Changes will take time to reach deployments

The projects discussed at KubeCon + CloudNativeCon North America 2026 are early pieces of a broader response to agentic AI. The source material does not identify a delivery schedule for their adoption across commercial Kubernetes distributions. It notes that the initiatives may take time to appear in the clusters IT teams operate.

For platform teams, the issues to watch are how agent sessions can be suspended and resumed without holding compute unnecessarily, whether control planes can handle higher API activity, and how runtime protections can constrain agent access. The current work indicates that Kubernetes remains extensible, while also underscoring that operating agent-heavy environments may require capabilities beyond those commonly used today.