1. News
  2. Cloud & infrastructure
  3. Google Cloud · AI agents
Google Cloud · AI agents3 min read

Google gives away three tools for running AI agents safely at scale

Google donated three open source tools for running AI agents to a neutral foundation, and says one of them fits 10x more agents on the same hardware.

By the Dashline deskSource: Google Cloud blog
Google gives away three tools for running AI agents safely at scale
Photo: Google Cloud blog

Google is giving away three of its own software projects built for running AI agents, the programs that take actions on their own. The tools are Agent Substrate, gVisor and OpenRL. Google donated them to the Cloud Native Computing Foundation (CNCF), a neutral group that oversees open source software. Anyone can now use them on any cloud.

What changed

The three projects were accepted into the CNCF Sandbox last month, which is the foundation's entry level for new projects. Google says each is now being set up with community maintainers and independent build systems. Agent Substrate also reached version 1.0 and is now generally available as a managed service on Google's GKE.

  • Agent Substrate: keeps agents from tying up computers while they wait.
  • gVisor: walls off code that an agent writes and runs.
  • OpenRL: helps teams train agents with reinforcement learning, using plain Python on Kubernetes.

Why agents need new plumbing

An agent works in a loop. It proposes an action, runs code or calls a tool, then feeds the result back to the AI model. Most of the time it is simply waiting on the model, another service or a person. Keeping a dedicated container running during those pauses wastes memory and computing power.

Agent Substrate saves an idle agent's memory and files to storage and frees up its computer. When the next request arrives, it restores the agent in under 500 milliseconds. Google compares it to a hotel room: you do not keep a room when you leave town, and a new one is ready when you return. Google says this allows 10x the density of standard containers and more than 500 suspend and resume actions per second.

Security is the other half. Agents often run unreviewed code and outside software packages. Standard containers share the host computer's core operating system, which Google calls a security risk. gVisor places that code inside a lightweight layer of its own. Agent Substrate also uses gateways that control what an agent can reach and keep sensitive keys out of the agent's environment.

Who it affects

The tools matter most to companies building or buying agent products. They run on any Kubernetes cluster, which is software that manages apps across many computers, or as a managed service on Google Cloud. The project has more than 100 contributors since public work began in May 2026. Nous Research is building Hermes Enterprise on it, and SPIRITT uses it for its agent workspace.

What is not known yet

The speed and density figures come from Google, and the material includes no independent tests. It also gives no pricing for Agent Substrate on GKE. gVisor is on a path toward the CNCF's Incubation level, but no timeline is given.

The takeaway

If your business plans to use AI agents, ask vendors how they pause idle agents and isolate the code those agents run.

More articles

You should really read these.

Next step

Let’s talk