Google said on 5 November 2026 that Agent Substrate, gVisor and OpenRL have been accepted into the CNCF Sandbox and are being donated to the Cloud Native Computing Foundation, in a Google Cloud Blog post and a companion Google Open Source Blog post. Google also said Agent Substrate has reached 1.0 and is generally available as Agent Substrate on GKE.
The performance figures in both posts are Google’s claims: restoring idle agents in under 500 milliseconds, running millions of sandboxes at 10x container density, and more than 500 suspend and resume activations per second. OpenRL is framed as Kubernetes-native orchestration for reinforcement-learning post-training. Sandbox application references cited in the brief are gVisor #521, Agent Substrate #523 and OpenRL #518.
In brief
- Agent Substrate, gVisor and OpenRL accepted into CNCF Sandbox; Google donating the projects.
- Agent Substrate 1.0 and GA as Agent Substrate on GKE (Google).
- Google claims: idle restore under 500 ms; millions of sandboxes at 10x density; >500 suspend/resume activations per second. OpenRL for K8s-native RL post-training.
Why agents break ordinary Kubernetes assumptions
Google’s thesis is that agents sit idle most of the time waiting on model turns, APIs or humans, yet still need stateful sandboxes and identity separate from a long-lived Pod. Agent Substrate is pitched as decoupling actor state from a warm worker pool so idle sessions can snapshot to storage and restore quickly. gVisor supplies userspace kernel isolation so untrusted agent-generated code does not talk straight to the host kernel. OpenRL aims to run the same sandbox environment used in production for RL rollouts on Kubernetes.
This is a new TSN post, not an update to the earlier Agent Compute Engine launch. ACE remains the managed microVM story; today’s CNCF move is about open governance for Substrate, gVisor and OpenRL plus Substrate 1.0 on GKE. Readers should keep the two threads linked but separate.
Density claims need outside clocks
Google’s 10x container density, sub-500-millisecond restore and 500 activations-per-second lines are the marketing spine of the announcement. They address a real pain — agents that hold Git state and runtimes while waiting on models — but they are still vendor measurements on Google’s stack. Platform engineers should ask for reproducible density tests on their own node shapes, including GPU and TPU workers, before rewriting capacity plans around Substrate 1.0 on GKE.
Open governance is the strategic half
Donating gVisor and Substrate to CNCF Sandbox is meant to clear single-vendor friction — Google’s Open Source post argues upstream Linux maintainers resisted host-kernel patches while gVisor was solely Google-owned. Whether Incubation follows is a CNCF process question, not a GKE marketing claim. For UK platform teams already on GKE Sandbox or evaluating agent density, the practical test is whether Substrate APIs stay portable across clouds once community maintainers and independent build infrastructure finish onboarding.
TSN’s coverage of other agent runtimes, from Gemini Omnimodal on the model side to Factory’s Droid harness claims on the coding-agent side, shows the same split: models and harnesses move fast, isolation and density move through infrastructure politics. Treat the 500 ms restore, 10x density and 500 activations-per-second numbers as Google’s published claims until independent benchmarks appear.
OpenRL is the newest and thinnest layer
OpenRL is described as the newest piece of the stack, brought to CNCF early on purpose so accelerator vendors and inference engines can share one RL recipe. That ambition is larger than a single Google blog post can prove. Until public recipes run on non-Google GPUs with published throughput, treat OpenRL’s Kubernetes-native pitch as directionally important infrastructure politics, not a finished training product buyers can standardise on tomorrow.
What to watch after KubeCon
Watch next for CNCF Sandbox onboarding completion, any path of gVisor toward Incubation, Substrate 1.0 release notes outside Google Cloud marketing, and OpenRL recipes that run on non-Google accelerators. Until independent density numbers land, Agent Substrate’s CNCF donation is a governance and GA announcement with Google-sourced performance claims — and a sibling story to ACE, not a rewrite of it.
ACE and Substrate stay separate stories
Readers who already bookmarked TSN’s Agent Compute Engine coverage should treat today’s CNCF donation as a sibling thread: managed microVMs on one side, open Substrate and gVisor governance on the other. Linking both posts helps, folding them into one URL would blur two different product surfaces Google is selling and donating at once.
Sources
- Powering open infrastructure for autonomous agents (Google Cloud Blog, 5 November 2026)
- Sandboxing Untrusted AI Agents at Scale (Google Open Source Blog, 5 November 2026)
Related stories
- Google Cloud Unveils Agent Compute Engine to Run AI Agents in MicroVMs
- Google Gemini Omnimodal Rolls Out
- Factory Says Droid 1.0 Beats Codex and Claude Code in Its Own Tests
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X





