VAST Data has announced an Agent Sandbox Foundation Stack, built with NVIDIA, that gives AI agents somewhere durable to keep their work: the stack pairs the VAST AI Operating System with NVIDIA’s NemoClaw blueprints and OpenShell runtime, so an agent’s state, memory and activity record persist even though the agent itself may exist for only a few minutes. The announcement was made in Singapore on October 22, alongside NVIDIA’s GTC Berlin.
The design starts from a problem production agent fleets create. Agents spawn sub-agents, execute tasks and disappear, and fleets can create and destroy thousands of isolated workspaces an hour. Long-running agents complicate things further: they work across sessions, sleeping, resuming and delegating to short-lived helpers. In both patterns, VAST argues, an agent’s progress, working data and memory must survive independently of any single sandbox or compute node — available, governed and auditable wherever execution resumes.
Ephemeral Compute, Durable State
The stack separates those two lifetimes. NVIDIA OpenShell provides the isolated execution environment and policy enforcement around agent activity; NemoClaw supplies NVIDIA’s open-source blueprints and reference stack for deploying specialised, always-on autonomous agents with governance controls. Underneath them, the VAST AI OS provides the persistent layer: rapid creation and removal of isolated agent workspaces, with zero-copy clones provisioning storage for new OpenShell sandboxes without duplicating the underlying data.
VAST DataStore provides persistent sandbox storage and snapshots, while VAST DataBase and its integrated data services give agents shared access to memory, context and enterprise data, so isolated agents can work concurrently from common data without redundant copies. Because execution and state are decoupled, an agent can pause, resume or move across eligible infrastructure without losing progress or context. The VAST Management System adds real-time visibility into data flows and infrastructure activity across the fleet.
A Record Of Everything The Agent Did
The governance half of the pitch is the audit trail. VAST DataBase maintains a persistent record of agent activity — tool calls, prompts, tokens, costs and decisions — that survives the agent that produced it. Alon Horev, VAST’s CTO and co-founder, put the case plainly: “Every enterprise that puts an agent into production will eventually need to answer what it did and why it did it. 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.”
Kari Briski, NVIDIA’s vice president of generative AI software, said that by combining NemoClaw and OpenShell with the VAST AI OS, “organisations can build agent sandbox environments that are secure and designed to scale from experimentation to production.” VAST says automated deployment of the stack will be available through a public GitHub repository, with demonstrations at GTC Berlin. The timing is deliberate: the EU AI Act’s record-keeping duties for high-risk systems make exactly this kind of logging a compliance requirement, not a nice-to-have.
Why It Matters
The VAST stack matters because it names the layer of the agent economy that model announcements skip: state. The industry’s agent safety conversation has concentrated on sandboxes and permissions — the ground GEroundtable covered with NVIDIA’s OpenShell reference architecture and CrowdStrike’s Falcon Guardian pairing with OpenShell — but a sandbox that forgets everything when it closes cannot answer an auditor, a customer or an incident investigator. VAST is betting that durable, queryable agent state becomes standard infrastructure, in the same way Grafana’s agent observability launch this week bet on watching agents from outside. Whoever owns the record of what agents did owns a piece of every regulated deployment, and storage vendors intend to be that owner.
Background
VAST Data, founded in 2016 and headquartered in New York, built its business on disaggregated storage for AI workloads and has been repositioning its platform as an “AI operating system” for the agent era. NVIDIA, for its part, has spent GTC Berlin turning OpenShell — its open, secure runtime for governing agent execution, access and inference routing — into an ecosystem: a reference architecture on its own hardware, security integrations such as CrowdStrike’s, and now infrastructure partners like VAST supplying the persistent data layer the runtime itself does not provide.
Source: Blocks & Files, Tech in Asia
0 Comments