Google is putting three of its agent-runtime projects under neutral cloud-native governance and shipping the centerpiece as a managed product. In a November 5, 2026 Google Cloud blog post, the company said Agent Substrate, gVisor and OpenRL have been donated to the Cloud Native Computing Foundation after those projects were accepted into the CNCF Sandbox last month. In the same announcement it said Agent Substrate has reached version 1.0 and is generally available as Agent Substrate on GKE.
Agent Substrate is the density layer. Google frames agent work as short bursts of code or tool calls separated by long waits on models, APIs or humans. Keeping a dedicated container alive through those pauses wastes memory and compute. Substrate splits the problem into Actors that hold memory, files and identity, Workers drawn from a warm Kubernetes pool, and Gateways that enforce access and inject secrets. When an agent pauses, its Actor is snapshotted to storage and the Worker goes back to the pool; when work resumes, Substrate restores the Actor onto any free Worker in under 500 milliseconds. Google says the design packs about ten times as many sandboxes as ordinary containers and has sustained more than 500 suspend and resume activations per second in testing, with more than 100 contributors since public development began in May 2026.
gVisor is the isolation layer. Agents often run unreviewed generated code, so a shared host kernel is a risky default. gVisor interposes a userspace application kernel so that code never talks directly to the host OS, while keeping a container-like memory model, GPU and TPU support, and checkpoint restore. Google open-sourced it in 2018; Anthropic, OpenAI, Modal, Snowflake, Ant Group and Tencent already run it at scale. Donating it to the CNCF is meant to unlock upstream Linux performance patches that maintainers previously declined while the project was single-vendor owned. Ant Group, Modal and Tines have committed to join maintainers, with OpenAI, Tencent and NVIDIA continuing contributions. A project blog dated October 2, 2026 had already outlined the donation timeline after CNCF accepted the application on September 28.
OpenRL is the training layer. It offers a self-hosted Python post-training loop on Kubernetes with four primitives (sample, forward_backward, optim_step, save_state) through a Tinker-compatible API, pluggable engines such as PyTorch AutoModel and vLLM, and DRA-based accelerator scheduling. Google pairs it with llm-d for prefix-aware rollout routing and GPU time-slicing between sampling and training. Gemma and Qwen are listed as validated starting points, with TPU support on the roadmap.
The managed path is Agent Substrate on GKE: the same runtime as a first-party capability on Google Cloud, or the open-source code on any Kubernetes cluster. Google positions it as complementary to Agent Compute Engine, its Titanium microVM service for hibernation and environment forking. Early partners cited include Nous Research building Hermes Enterprise on Substrate, SPIRITT running agent workspaces on GKE, and other CNCF projects such as kagent.
For builders of coding agents, eval fleets and RL loops, the news is less a new model than a bet that agent sandboxes should look like Kubernetes did for containers: open, portable and dense enough to be cheap. Claims about density, restore latency and contributor counts come from Google's own post and should be read as vendor figures until independent operators publish their own numbers.