Verified, then logged.
Sun 22 Nov 2026 · 00:22 KST / 15:22 UTC한국어 | ENSearch /

02 / IT · Analysis

Google ships Agent Substrate 1.0 on GKE and gives it to CNCF

Agent Substrate moves to the CNCF alongside gVisor and OpenRL, but public documents give no test conditions for its density figures and no list of breaking changes.

In a data center aisle, a rack of mostly idle servers with unlit status lights stands beside a storage rack with flickering drive lights, while a technician is seen from behind, out of focus, in the distance. AI-generated image

Google announced general availability of its agent runtime Agent Substrate 1.0 on Google Kubernetes Engine (GKE), leaving teams that run agents on Kubernetes to choose between a managed capability and their own deployment. In the same announcement, Google said it is donating the project, along with gVisor and OpenRL, to the Cloud Native Computing Foundation (CNCF).

Google's announcement ties the managed product's launch to the open-source projects' move to the foundation. The GKE managed capability runs on Google Cloud, while the open-source projects go to CNCF.

The density multiple for the same runtime, however, differs between Google's announcement and the CNCF application.

How three projects move to the CNCF

The three projects were accepted last month into the CNCF Sandbox, the foundation's entry-level tier for early-stage projects, Google said. Agent Substrate's application, issue #523, was opened on 8 Sep 2026 and has been closed with a label showing that the vote passed. Google's announcement does not include vote records for the other two projects.

gVisor is an isolation kernel released as open source in 2018. According to Google, it handles system calls in user space, so untrusted code does not communicate directly with the host kernel. Ant Group, Modal and Tines have committed to joining its maintainers.

The foundation move is also tied to host kernel patches that raise gVisor's input/output performance. Google said upstream Linux maintainers declined to merge them while the project had a single vendor owner, and it argues that the move to the foundation opens that route.

Google describes OpenRL as a tool that runs reinforcement-learning post-training loops in Python on Kubernetes. Post-training refines a model after its initial training. The company wrote that "OpenRL is the newest layer of this stack by a wide margin", and it is moving the project to neutral governance before the design hardens. OpenRL offers a Tinker-compatible API, and existing tinker-cookbook recipes run without changes.

A design that sends idle agents to storage

Standard Kubernetes gives each agent its own Pod, the platform's smallest deployable unit. Google said this ties up memory and CPU even while the agent is waiting. The number of objects tracked by the control plane, the cluster's management layer, grows as well.

According to Google, Agent Substrate splits into three components. An Actor holds an agent's memory, files and identity. A Worker is a reusable execution sandbox. A Gateway handles access control and injects secrets. When an agent is idle, its state is snapshotted to storage. The Worker returns to a reusable pool in the meantime.

Google's announcement puts density at 10x that of standard container runtimes. The CNCF application puts density at 30x to 100x through worker multiplexing, which shares a small pool of workers among many agents, but it states no baseline for that comparison. Neither document explains the gap or names the hardware used for measurement.

How adopting teams choose to run it

Deployment options. According to Google, GKE users run the software as a managed capability on Google Cloud. Teams that operate it themselves deploy the open-source code to a Kubernetes cluster. The isolation boundary is either gVisor or a Cloud Hypervisor microVM, a lightweight virtual machine. The announcement does not give pricing or a service-level commitment.

Compatibility. Google said changing the isolation boundary leaves the way agent lifecycles are managed unchanged. The announcement does not list the breaking changes in 1.0, meaning which APIs or settings have changed.

Licence and commercial terms. The CNCF application lists Agent Substrate's licence as Apache 2.0. It also says the project is not a commercial product or SaaS offering, and it includes an agreement to donate trademarks and accounts to CNCF if the project is accepted. Google's announcement says the GKE managed capability runs separately on Google Cloud. The announcement does not give licences for gVisor or OpenRL.

A deployment range that reaches local clusters

Where it runs. According to the application, Agent Substrate also runs on kind local clusters and on-premises, as well as on cloud-managed Kubernetes. kind is a tool for running Kubernetes clusters on a local machine. Google said gVisor deploys to Linux nodes without nested virtualisation. OpenRL has been validated with Gemma and Qwen, and TPU support is on its roadmap.

Cryptographic identity and mutual TLS are on the application's roadmap, but Google's announcement does not say how much of that work is in 1.0. On the gVisor kernel patches, Google's account goes as far as saying the route is open. The decision on accepting them rests with Linux maintainers and has not been made.