Skip to navigation

Secure Agent Workspace

View as Markdown
View Secure Agent Workspace in the interactive architecture

Secure Agent Workspace is a Capability Reference Design that adds a governed environment for persistent and autonomous agent work. A user endpoint attaches to the workspace, while identity, network access, credentials, runtime policy, audit, and human review constrain execution outside the model trust boundary.

The published Building secure agent workspaces with NVIDIA OpenShell Reference Architecture describes an OpenShell-based composition of this capability. Use this page to understand the capability’s relationship to the core. Refer to the published architecture for detailed design and implementation guidance.

Inference is a dependency, not part of the workspace boundary. A workspace can consume a routed model endpoint or local inference without turning the inference service into a workspace component.

Relationship to the Core

Secure Agent Workspace extends the core layers as follows:

Core LayerExtension
WorkloadsAdds persistent agent state, tool authority, long-running execution, delegation, and human-review requirements.
AI platformInherits model lifecycle and serving through governed interfaces. Routed inference remains outside the workspace boundary.
Platform servicesAdds credential and authority mediation plus network perimeter controls. Extends access brokering, audit, observability, and AgentOps.
Cloud-Native PlatformAdds a policy-selected isolation boundary and extends workspace lifecycle, runtime policy, storage, attachment, and fleet operations.
Hardware foundationInherits the Enterprise Hardware Reference Architecture and consumes the virtualization or isolation capabilities required by the selected workspace pattern.

Architecture

Managed Workspace
A single-user Linux or Windows environment with persistent files, tools, processes, agent state, and optional GPU acceleration.
Lifecycle Control
An authoritative portal or API for create, start, stop, expose, connect, suspend, revoke, and decommission operations.
Runtime-Bounded Execution
A sandbox and policy layer that mediates files, processes, network destinations, credentials, devices, and tool calls.
Brokered Enterprise Access
Enterprise SSO, bounded sessions, short-lived credentials, controlled connectors, and attributable actions.
Network Boundary
Deny-by-default egress, approved destinations, routed inference, internal service access, and observable policy decisions.
Governance and Audit
Delegation records, policy, telemetry, human-review gates, revocation, investigation, and fleet operations.

NVIDIA OpenShell is an implementation option for runtime-bounded agent execution. It provides sandbox lifecycle, policy controls, credential mediation, network controls, and multiple compute drivers. OpenShell does not replace the selected container sandbox, microVM, VM, cluster network policy, or enterprise governance boundary.

NVIDIA Open Agent Safety Platform is related guidance for full-stack agent governance and continuous monitoring. Secure Agent Workspace composes workspace lifecycle, identity, isolation, network, credential, and audit requirements; Open Agent Safety Platform does not replace that workspace boundary. Review the technical architecture overview and OpenShell Developer Guide when selecting implementation controls.

Secure Agent Workspace requires an approved agent isolation boundary. A raw pod, container, or shared host process without a policy-approved sandbox runtime or VM boundary does not satisfy this capability.

Required Interfaces and Dependencies

  • Enterprise identity and accountable user or sponsor.
  • Workspace identity, lifecycle state, and attachment controls.
  • Logical agent registration and per-engagement delegated authority.
  • Short-lived, runtime-bound credentials for individual tool actions.
  • Routed inference or explicitly governed local inference.
  • A policy-approved isolation class, such as a microVM-backed Kata Containers runtime or a dedicated VM provided through KubeVirt or OpenShift Virtualization.
  • Default-deny network policy, controlled ingress, and policy-routed egress through an egress gateway or equivalent enforcement point.
  • Auditable connectors to enterprise data and actions.
  • Fleet health, security operations, incident containment, and decommissioning.

Security and Governance Boundary

Keep user, workspace, logical-agent, runtime-agent, and tool-call identities distinct. Bind each action to the sponsoring user and the policy state that authorized it. Prompt instructions are not a substitute for enforceable runtime, credential, and network controls.

Non-Goals

  • An enterprise identity provider.
  • A universal inference service.
  • A replacement for endpoint security or source-system authorization.
  • One prescribed cloud, hypervisor, Kubernetes distribution, or agent framework.

Deployment Patterns

Select a deployment pattern that satisfies the workload’s isolation, operations, and acceleration requirements:

PatternTypical Use
Container-sandboxed workspaceKubernetes-integrated workspaces using a policy-approved container sandbox when its isolation properties satisfy the workload risk profile. Refer to OpenShell runtime options.
MicroVM-isolated workspaceKubernetes-scheduled workspaces using a supported microVM-backed runtime configuration, such as Kata Containers, to provide a hardware-virtualized workload boundary while preserving pod-oriented lifecycle and policy integration. Refer to Kata Containers documentation.
VM-isolated workspaceDedicated virtual machines using KubeVirt or OpenShift Virtualization when stronger workload separation or guest operating system control is required. Refer to Published OpenShift Virtualization example.
CPU workspace with routed inferenceStarter deployments, coding, research, and documentation workloads. Refer to Inference Hub integration.
GPU-enabled workspaceAccelerated tools, local experiments, or bounded local inference. Refer to GPU Operator documentation.
Shared platform with per-user workspacesCentral capacity with workspace and tenant isolation. Refer to Complete SAW reference design.
Centrally managed fleetEnterprise rollout with standard policy, lifecycle, audit, and security operations. Refer to Complete SAW reference design.

The selected runtime class, hypervisor, host configuration, device exposure, and attestation model must be supported and approved for the applicable risk tier. Naming a runtime alone does not establish the required isolation boundary.

Operations

Operate workspace images, state, policy, credentials, identity bindings, network access, connectors, audit, capacity, backup, incident containment, revocation, and secure decommissioning as one lifecycle.

Published Reference Implementations

Implementation links describe specific partner-platform patterns. Their inclusion does not make either platform the universal architecture.

Primary Source

Building secure agent workspaces with NVIDIA OpenShell Reference Architecture