Google announced on November 5, 2026 that it is donating three open source projects to the Cloud Native Computing Foundation (CNCF): Agent Substrate, gVisor and OpenRL. The projects were admitted to the CNCF Sandbox and now move to neutral, vendor-independent governance. Alongside the donation, Agent Substrate reached version 1.0 and became available as a managed service on Google Kubernetes Engine (GKE).
Why AI agents need different infrastructure
According to Google, AI agents work differently from traditional applications: they run untrusted code, commands and tool calls, and spend 90% to 95% of their time idle, waiting for model inference, external APIs or human review — yet they cannot be shut down, because they hold state in memory and on disk.
Keeping a dedicated Pod per agent running all day solves latency, but the cost becomes unsustainable at scale. Traditional serverless saves compute, but cold resumes take more than 20 seconds. Agent Substrate was built for the middle ground: stateful actors that wake in under 500 milliseconds, without paying for idle Pods.
The three layers of the stack
- gVisor (secure isolation): a user-space application kernel written in Go that intercepts system calls and isolates untrusted code from the host kernel, with fast memory checkpointing and restore.
- Agent Substrate (agent runtime): an actor-based control and data plane on top of Kubernetes. Actors hold session state; Workers — pre-warmed Pods in a shared pool — provide the compute. Thousands of agents can activate, suspend and resume every second without creating Pod objects or overloading etcd.
- OpenRL (training orchestration): a self-hosted framework for reinforcement-learning post-training, written as plain Python loops on Kubernetes. It began as a research preview in February 2026 and was donated early on purpose, according to Google.
Google also points to llm-d, a distributed inference project already in the CNCF that recently reached version 1.0, as a companion piece for OpenRL workloads.
Agent Substrate is a different layer from the Agent Compute Engine, launched by Google Cloud in October: while the managed service gives each agent its own microVM, Substrate is the open source runtime that organizes agents on top of Kubernetes.
Neutral governance and next steps
With the donation, the cloud-native community gains a neutral foundation that works on any conformant Kubernetes cluster. According to the gVisor blog, cited by TechTimes, Ant Group, Modal and Tines have signed on as committed maintainers, and OpenAI, Tencent and NVIDIA continue as contributors.
Google is taking the stack to KubeCon + CloudNativeCon North America in Salt Lake City, November 9–12, with a full day dedicated to Agent Substrate — ATEcon — on Monday, November 9.
Caveats: known limits and a 1.0 still stabilizing
Not everything is settled. TechTimes notes that Substrate's own documentation lists limits: there is no GPU passthrough for actor containers, and gVisor cannot snapshot live CUDA contexts. The outlet also flags a discrepancy: while Google announces version 1.0, the project's GitHub README still described it as pre-1.0, with a v1.0.0-rc.1 release candidate available. Sandbox admission dates also vary by source: the gVisor blog records acceptance on September 28, while Google places the admissions in October.
Image credit: Google Open Source Blog.
Sources: Google Open Source Blog; Vinish.Dev; TechTimes.
Transparency: this article was produced with the help of artificial intelligence, based on cross-checking public posts and sources on the internet, including the sources cited above. We recommend consulting the original sources to confirm the details.
By Muse
.jpg)
.jpg)
