Google announces two open source projects for Kubernetes: AI Conformance (ensuring AI workloads have portable, standardized cluster requirements) and Agent Sandbox (providing secure isolation for running autonomous AI agents that can execute code and commands). Both projects aim to make Kubernetes the best platform for AI training, inference, and agentic workloads.
Gary Singh, Product Manager at Google, discusses two major announcements for the Kubernetes and CNCF community made at KubeCon.
The first announcement is AI Conformance for Kubernetes. Just as Kubernetes conformance ensures workload portability across platforms, AI Conformance extends this to AI-specific workloads. This allows customers to confidently choose between different vendors knowing their AI training and inference workloads will work consistently.
The goal is to establish standards that make AI a more official part of the CNCF and Kubernetes ecosystem, giving users confidence in workload portability.
The second announcement is Agent Sandbox, addressing the critical need for secure isolation when running agentic AI. As autonomous agents can run commands and generate code, organizations need ways to sandbox this activity within Kubernetes.
Agent Sandbox brings together various isolation technologies (like gVisor) into an integrated solution with new Custom Resource Definitions (CRDs). This makes it easier to host and run agents securely while still benefiting from agentic operations and code generation.
Both projects are open source under Kubernetes. AI Conformance has no additional charge—it benefits the community by ensuring conformant clusters. For Agent Sandbox, pricing is simply based on compute usage on Kubernetes; Google is providing better tooling for running agents, not a premium service.
"How do we sandbox and isolate this code to ensure that it doesn't—technical term—wreak havoc on your environment?"
"This starts to make Kubernetes the best place to not just run AI training and inference but now to actually run agentic workloads as well."
"We're just giving you a better way to run agents."