VAST Data and NVIDIA have announced an “Agent Sandbox Foundation Stack” they say is designed to help enterprises run large numbers of AI agents and sub-agents in production while keeping records of agent activity for governance and review.
According to the companies, the stack is intended to provide persistent data infrastructure for “thousands of AI agents and sub-agents” running in isolated sandboxes, with a focus on retaining operational records after agents terminate. VAST said the system maintains a persistent record of agent tool calls, prompts, tokens, costs and decisions.
“Every enterprise that puts an agent into production will eventually need to answer what it did and why it did it,” said VAST CTO and co-founder Alon Horev, in a statement. NVIDIA’s vice president of generative AI software, Kari Briski, said the goal was to support secure execution while preserving “data, context and state” over the lifecycle of agent workloads.
The announcement lands as organisations look to operationalise “agentic” AI systems that can autonomously execute tasks, spawn other agents and operate across multiple sessions. The companies argue that these patterns create a need to decouple persistent state from short-lived compute and sandbox environments so agents can pause, resume or move without losing context.
In Australia, the development also follows guidance from the Australian Signals Directorate on “agentic AI harnesses” for organisations adopting autonomous agents. In the email accompanying the release, the sender said the ASD guidance calls for system isolation, continuous monitoring and human approval checkpoints for agentic AI.
VAST said the stack combines its “AI Operating System” with NVIDIA NemoClaw and NVIDIA OpenShell. The release describes OpenShell as providing isolated execution and policy enforcement around agent activity, while VAST provides persistent state, scalable workspaces and operational records across the agent lifecycle.
The companies outlined three layers in the architecture: infrastructure for dynamic agent environments, persistent state and integrated data services, and observability and auditability. VAST said its platform uses “zero-copy clones” to provision persistent storage for isolated sandboxes, and that its systems maintain records intended to serve as an audit trail across multi-agent workflows.
VAST said it will demonstrate the stack at NVIDIA GTC Berlin and plans to make an automated deployment available through a public GitHub repository.
