Rather than rushing AI agents into production, large enterprises, tech companies, and other security-conscious organizations are keeping agents on a tight leash until they can get them under control.
When we set AI agents free to solve a problem or complete a task without first determining every step they should take, we’re treating them like human employees. Unlike human employees, however, AI agents lack accountability. There’s no concern of repercussions if their actions negatively impact the business. All you can do is attempt to unwind any damage and ensure the same mistake doesn’t happen again.
Also unlike humans, agents can proliferate as fast as we want them to. One agent today could become 1,000 agents tomorrow, all without HR lifting a finger.
Agent sandboxes are quickly becoming the de facto approach for implementing guardrails around AI agents. Sandboxes should prevent agents from going rogue and limit the blast radius of any questionable actions, while still providing agents access to all the tools they need.
Introducing the VAST Foundation Stack for Governed Agents
As a result of our continued collaboration with NVIDIA, our customers now have a blueprint for deploying fully capable agents inside highly secure sandboxes: the VAST Foundation Stack for Governed Agents. This architecture combines the VAST AI OS with the NVIDIA Agent Toolkit, including NVIDIA OpenShell secure runtime, and NVIDIA NemoClaw blueprints for building and deploying agents. It also includes NVIDIA Sentry for hardware-based security enforcement.
Each piece plays an integral role in the lifecycle and governance of an AI agent:
NVIDIA NemoClaw provides the blueprints for building, deploying, and operating functional agents across a wide range of enterprise needs. It supports Hermes, LangChain, OpenClaw, or custom agent harnesses.
NVIDIA OpenShell provides a secure runtime environment for these agents. It uses an identity-based approach to enforce access and behavior policies, and it logs every interaction between an agent and the world outside its sandbox.
NVIDIA Sentry is an additional layer of defense for organizations deploying the NVIDIA Vera Rubin platform. It utilizes the onboard NVIDIA BlueField-4 DPUs to provide real-time monitoring and security between agents running on NVIDIA Vera CPUs and models running on NVIDIA Rubin GPUs, as well as other external systems.
VAST AI OS delivers the infrastructure to run AI agents at scale and extend security controls into the data layer, focusing on observability/auditing, AI-native storage capabilities, and operational data services.
Here’s a more detailed look at the jobs each component performs during the AI-agent lifecycle:
1. NVIDIA OpenShell establishes the security rules 2. NVIDIA OpenShell provides kernel-level sandbox isolation 3. VAST AI OS powers the underlying operational layer 4-6. The agent harness running in OpenShell triggers new agent tasks 7. VAST DataStore’s Block CSI uses zero-copy clones to provision OpenShell sandboxes in seconds 8. VAST DataBase serves as the shared vector store and knowledge base, allowing hundreds of reasoning agents to query the exact same dataset concurrently without data sprawl 9. VAST DataStore provides isolated, persistent volumes that allow agents to suspend and resume using snapshots and clones, as well as dedicated workspaces and shared storage spaces for objects and files 10. VAST Management System delivers real-time data-layer telemetry that seamlessly interoperates with NVIDIA Sentry 11. When NVIDIA Sentry flags or quarantines a misbehaving agent, VAST AI OS instantly supports those policies by isolating data access and freezing the sandbox state for immediate forensic review 12. OpenShell enforces governance, and the VAST DataBase records logs every tool call, prompt, token, cost, and decision for an immutable audit trail
Powering agents with high-performance data services
Once identities are assigned, policies are defined, and governance infrastructure is in place, AI agents need to do real work on real data.
The VAST AI OS lets agents access and execute on permitted data in real-time, whether that’s raw data stored as objects or files in DataStore, or tables and vectors in VAST DataBase. VAST DataEngine, SyncEngine, and EventBroker enable full data pipelines (including real-time ETL, multi-modal analytics, and long-term, high-fidelity data retention), allowing agents to execute compute functions or entire inference workflows as new data is ingested.
A demonstration of governed agents in action
To get a better sense of how this looks behind the scenes from the VAST AI OS perspective—including sandbox creation, event logging, and human-in-the-loop approvals—watch this demonstration involving a fleet of agents deployed to assess strategic business opportunities.
Enterprise AI without compromising security and compliance
The transition from experimental AI to production level agents shouldn’t require compromising on enterprise security or operational control. As long as agents operate in the dark—without strict boundaries, robust memory, and granular data governance—they will be a liability rather than a scalable business asset.
The VAST Foundation Stack for Governed Agents with NVIDIA NemoClaw fundamentally changes this dynamic. By uniting the secure agentic runtime and policy enforcement of NVIDIA OpenShell and Sentry with the highly scalable, zero-trust infrastructure of the VAST AI OS, organizations now have a comprehensive, deployable blueprint to bring agents into production safely. This architecture bridges the critical gap between powerful AI capabilities and strict enterprise data governance.
This foundation stack builds on the recently announced VAST DataEnclave, which delivers a trusted and cryptographically secured runtime for running leading commercial AI models alongside sensitive data on customer-controlled infrastructure. By injecting trust and control into every layer of AI operations—data, models, and agents—we’re making it easier for security-conscious organizations to reap the benefits of AI innovation.
The future of enterprise AI isn’t just about building smarter agents; it’s about building trust. With VAST and NVIDIA, the foundation for that trust is finally here.
Preview the VAST Foundation Stack for Governed Agents here.
