Google unveils new open source tools and Kubernetes features to make AI workloads faster, safer, and easier to manage, strengthening its role in the open AI ecosystem.
At KubeCon + CloudNativeCon North America 2025 in Atlanta, Google LLC unveiled three major innovations aimed at simplifying and accelerating AI operations, two open source tools, Agent Sandbox and Multi-Tier Checkpointing (MTC), along with a new Google Cloud feature, GKE Pod Snapshots. Together, these solutions enhance the performance, security, and resilience of AI workloads running on Kubernetes.
Agent Sandbox, built as a Kubernetes extension, enables developers to dynamically create and delete thousands of isolated AI agent environments. Designed to address cybersecurity risks from AI agents interacting with external applications, it is built on gVisor, Google’s open source container isolation technology first released in 2018. The integration with Google Kubernetes Engine (GKE) allows the creation of pre-warmed sandbox environments, ready for immediate use and optimised for faster processing.
Complementing this, GKE Pod Snapshots dramatically reduces the startup times of large language models, which typically take up to 10 minutes to initialise. The feature stores ready-to-use container snapshots, including configurations, reducing startup time by as much as 80%.
“GKE Pod Snapshots supports snapshot and restore of both CPU- and GPU-based workloads, bringing pod start times from minutes down to seconds,” said Brandon Royal, Senior Product Manager, Google. “With Pod Snapshots, any idle sandbox can be snapshotted and suspended, saving significant compute cycles with little to no disruption for end-users.”
The second open-source tool, Multi-Tier Checkpointing (MTC), streamlines AI training by speeding up recovery after errors, enabling faster retraining and updates. Collectively, these tools reinforce Google’s commitment to open, scalable AI infrastructure that empowers developers without binding them to proprietary frameworks.
