What changed
Google LLC has launched Agent Substrate, an open-source project built on Kubernetes, aimed at securely isolating AI agents. The project was donated to the Cloud Native Computing Foundation alongside its gVisor sandbox. The initiative addresses the growing need for secure and efficient execution of agentic workloads in enterprise environments.
What this means in practice
Agent Substrate provides hardened sandbox runtimes densely packed on Kubernetes clusters. All traffic passes through a policy layer that understands identity, ensuring agents can only access resources explicitly granted. Credentials are kept out of the agent’s reach by injecting them on the wire, reducing exposure risks.
The project also snapshots idle agents out of memory and CPU, storing them offline and restoring them when needed, which frees up resources across the cluster. This helps manage compute costs during periods of inactivity, such as waiting for a large language model response or a human follow-up.
Google Cloud’s product manager, Brandon Royal, emphasized that agents must be fully isolated, whether they are accessing tools, using a computer, or driving a browser. This isolation is critical for security and operational efficiency.
Limitations and availability
Agent Substrate is currently available as an open-source project, with Gemini Managed Agents already running on it. Google Cloud has also introduced Agent Compute Engine to host AI agents. While the project is in active development, it is not yet universally available across all platforms or configurations. The supplied announcement does not specify exact deployment timelines or hardware requirements for running Agent Substrate.
