VAST & Nvidia launch AI agent sandbox for enterprises

VAST Data and Nvidia have launched an AI agent sandbox foundation stack for enterprise deployments. The offering is designed to support the secure operation of thousands of AI agents and sub-agents in production.

The stack combines the VAST AI Operating System with Nvidia NemoClaw and Nvidia OpenShell. It is intended to keep persistent agent state separate from short-lived compute environments, allowing agents to pause, resume or move between eligible infrastructure without losing progress or context.

That addresses a growing problem as companies test more autonomous software agents inside business systems. Agents may create other agents, complete tasks over multiple sessions and then disappear, leaving organisations needing to retain records of what happened inside each isolated workspace.

VAST said its system keeps a persistent record of agent activity, including tool calls, prompts, tokens, costs and decisions. Those records remain available after the agent has ended, giving platform teams an audit trail across multi-agent workflows.

Under the architecture, Nvidia OpenShell provides isolated execution and policy enforcement, while VAST supplies persistent state, data services and operational records. According to the companies, this avoids tying long-term agent memory and working data to any single compute node or sandbox.

Three layers

VAST described the stack as spanning three main layers. The first is infrastructure for dynamic agent environments, where isolated workspaces can be created and removed quickly as tasks begin and end.

In this layer, VAST uses zero-copy clones to provision persistent storage for OpenShell sandboxes without making full copies of the underlying data. The aim is to support workflows in which agents can be created, disappear and reappear as reasoning tasks change.

The second layer covers persistent state and data services. VAST said its DataStore product handles persistent sandbox storage and snapshots, while VAST DataBase and related services provide access to agent memory, context and enterprise data so separate agents can work from shared information without duplicating it.

The third layer focuses on observability and auditability. VAST said its management system provides real-time visibility into infrastructure activity across fleets of agents, while VAST DataBase stores a durable record of the actions those agents take.

Alon Horev, Chief Technology Officer and Co-Founder at VAST Data, said the system is aimed at a governance gap businesses are likely to face as agent use expands.

"Every enterprise that puts an agent into production will eventually need to answer what it did and why it did it," said Alon Horev, Chief Technology Officer and Co-Founder at VAST Data.

"An agent might exist for minutes, spawn thousands of other agents and then disappear, but its work has to persist and its actions have to stay understandable long afterward. Together with Nvidia, we're showing how the VAST AI Operating System provides the durable foundation underneath secure agent environments, so organisations can scale agentic systems without sacrificing persistence, governance or control," Horev said.

Nvidia framed the partnership around the practical demands of enterprise deployment, particularly where agents operate across several sessions and interact with shared business data.

"As AI agents take on more complex work, enterprises need infrastructure that can support secure execution while preserving the data, context and state agents depend on across their lifecycle," said Kari Briski, Vice President of Generative AI Software at Nvidia.

"By combining Nvidia NemoClaw and OpenShell with the VAST AI Operating System, organisations can build agent sandbox environments that are secure and designed to scale from experimentation to production," Briski said.

Broader push

The launch extends VAST's broader work around what it calls foundation stacks: packaged implementations intended to help organisations move AI projects into repeatable production environments. Earlier stacks were built around enterprise research assistance and video intelligence use cases using Nvidia models and software.

This latest version shifts the emphasis from application examples to the underlying infrastructure needed to run autonomous agents. The central argument from both companies is that agent execution can remain temporary while the agent's state, records and working data persist independently.

That distinction is likely to matter for companies trying to balance automation with governance requirements. In practice, businesses adopting AI agents will need ways to review past actions, trace decisions and manage costs across large numbers of short-lived software processes.

VAST said the stack is meant to give organisations a repeatable architecture for doing that while maintaining isolated workspaces for agents and sub-agents. The persistent record includes tool calls, prompts, tokens, costs and decisions.