Google is handing the plumbing of the AI agent era to the open-source world. On November 5, the company said it will donate Agent Substrate, the runtime it built to run AI agents at scale, and gVisor, its sandboxing kernel, to the Cloud Native Computing Foundation (CNCF), in Sandbox status, alongside a training project called OpenRL. The move lands just before KubeCon North America, which runs November 9 to 12 in Salt Lake City, and it comes with a 1.0 release of Agent Substrate for Google Kubernetes Engine.
What Google handed over
The donation covers three projects Google presents as one open stack for agentic infrastructure. The first layer is gVisor, the user-space kernel that isolates untrusted code from the host machine. The second is Agent Substrate, the runtime that starts, parks and resumes agent sessions on shared compute. The third is OpenRL, which coordinates reinforcement-learning loops and evaluation sandboxes on Kubernetes.
Google’s stated reason is straightforward: without a shared open standard, every organisation ends up building its own custom scheduler on top of virtual machines, or working around Kubernetes. Donating the projects to the foundation gives the community what Google calls a neutral, vendor-independent foundation that works on any conformant Kubernetes cluster, not just Google’s cloud.
Why idle agents break Kubernetes
The problem Substrate attacks is idle time. Google says an agent in an interactive session spends 90 to 95 percent of its time waiting, for a model to stream tokens, for an external service to answer, or for a human to review a pull request. While it waits, the agent uses no processor power, yet its state cannot be discarded.
Platform teams face two poor options. Keep a Pod running all day and the agent answers instantly, but the bill grows with every idle hour, and Google says costs become unsustainable at around ten thousand users. Scale to zero and the bill shrinks, but waking the agent means cloning repositories, reinstalling packages and rebuilding memory, pushing delays past twenty seconds.
Training and evaluation hit the opposite wall. Reinforcement-learning loops and benchmark suites need thousands of short-lived sandboxes at once, and creating and destroying that many Pods every minute overwhelms the control plane. Google says the churn saturates the API server, overloads its database, and exhausts subnet IP addresses.
How Substrate works
Substrate splits an agent’s state from the machine that runs it. An Actor holds one session’s state: its filesystem, environment variables, memory and network rules. Workers are pre-warmed Kubernetes Pods in a shared pool, each equipped with container runtimes and sandbox boundaries, ready to host whichever Actor needs compute next.
When a request arrives for an Actor, a proxy called atenet-router holds it in memory and alerts the controller. The controller places the Actor in a slot in the warm Worker pool, restores its memory and state from tiered storage, and delivers the request. Google’s target is a resume in under 500 milliseconds, with the agent continuing where it paused.
Security is built into the design. Agents run only inside gVisor or Cloud Hypervisor microVM sandboxes, never in Linux containers, because standard containers share the host kernel. Networking is default-deny, Actors hold no credentials, and an egress gateway injects them into approved outbound requests. An agent that escaped its sandbox and robbed a real company shows why that isolation matters.
This is a different layer from Google’s managed Agent Compute Engine, which gives every Gemini agent its own isolated microVM as a Google Cloud service. Substrate is the open-source runtime underneath that kind of product, available to run on any Kubernetes cluster.
Our read
The donation is a standards play, and Google is making it from a position of strength. Kubernetes won the container era because Google gave the orchestrator away and let the industry standardise on it. Substrate is an attempt to repeat that trick for agents: define the runtime, give it to a neutral foundation, and let others build on Google’s terms.
That only works if others adopt it. The project is framework-agnostic and runs standard container images, which helps, and Solo.io has already added Substrate to its kagent tool as a selectable runtime. But the maturity question is real, and it cuts the other way.
The 1.0 label itself checks out. The project’s README says Agent Substrate has reached 1.0 with a commitment to API compatibility, and GitHub lists a published v1.0.0 release after three release candidates. The release page names it ‘Fully armed and operational.’
The caution is different: it is a first major release, the donation enters the CNCF at Sandbox status, the foundation’s entry tier, and the project’s own contribution guide still calls the effort very young. Teams should treat it as promising infrastructure to test, not as a settled production platform yet.
The timing is also deliberate. KubeCon opens November 9 with a full Agent Substrate Day, a hands-on workshop on November 11, and a live agent-swarm demo on the show floor. Google is not quietly donating a side project; it is putting the runtime at the centre of the cloud-native world’s biggest week.
Why It Matters
If agents become the default way software gets built and run, the runtime underneath them becomes as important as the models themselves. Whoever defines how agents are isolated, parked and woken up shapes their cost, their security and their portability. An open standard lowers the lock-in risk for every team betting on agents.
The research shows the stakes. Nine AI agents tried research work and none passed, which shows how far agents have to go before they can be trusted with tasks. Better runtimes will not fix weak reasoning, but they remove one excuse: agents that are cheap to keep, fast to wake and safely sandboxed are easier to test, evaluate and improve.
Background
Google introduced Kubernetes twelve years ago, and Substrate is its answer to a workload Kubernetes was never designed for. The project was introduced in May 2026 with GKE Agent Sandbox, which Google says grew more than sixteenfold in under five months. Days earlier, Google open-sourced the engine behind its products, the XLA system behind Search, YouTube and parts of Gemini.
Source: Google’s announcement and Daily TechNow’s report
0 Comments