VAST Data, NVIDIA launch Agent Sandbox Foundation Stack

The announcement extends a partnership that already produced earlier Foundation Stacks — one for enterprise research assistants built on NVIDIA Nemotron, and one for video intelligence using NVIDIA Cosmos inference models and NVIDIA Video Search and Summarization. Those earlier stacks turned NVIDIA AI Blueprints into reproducible, production-ready implementations on VAST AI OS; the new one shifts the focus from AI applications to the agents themselves.

The problem the stack addresses is structural. As VAST describes it, agents spawn other agents, create and discard thousands of isolated workspaces an hour, and long-running agents may work across multiple sessions, pausing and delegating to short-lived sub-agents. Progress, work data and memory therefore need to persist independently of any single sandbox or compute node — available, governed and auditable wherever execution resumes.

The division of labor is explicit: NVIDIA OpenShell isolates agent execution and applies policies to its activity, while VAST supplies persistent state, scalable agent workspaces, integrated data services and operational records. VAST Management System adds real-time visibility into data flows and infrastructure activity, and VAST DataBase retains records of tool calls, prompts, tokens, costs and decisions — the basis for a persistent audit trail across multi-agent workflows. Whether enterprises adopt it at scale may hinge on how the promised GitHub deployment and GTC Berlin demonstration land.