Skip to main content

Concepts

Understanding AgentVisor™'s core concepts helps you build secure, reliable agents.

Overview​

AgentVisor combines several technologies to provide secure agent execution:

  • Sandbox Isolation: gVisor or container-based isolation
  • Policy Enforcement: Fine-grained access control via Manetu PolicyEngine
  • Durable Execution: Temporal workflows for reliability
  • Transparent Proxying: HTTP requests intercepted and policy-checked

Key Concepts​

Architecture​

How the host and guest runtimes communicate, and the security boundaries between them.

Sandbox Modes​

The three isolation levels: none, docker, and gvisor.

Transports​

How external clients access AgentVisor agents through protocol-specific API endpoints (OpenAPI, MCP, A2A).

Gateways​

How agents access external services (MCP servers, other agents) through policy-enforced, credential-protected connections.

Credential Brokering​

How API keys are protected using symbolic tokens, ephemeral TLS termination, and host-side credential substitution.

HTTP Proxy​

How HTTP requests are transparently proxied through the host runtime.

Egress Controls​

How the policy gate and the always-on SSRF guard independently control which outbound destinations an agent can reach.

Temporal Integration​

How Temporal provides durable execution and state management.

Policy Enforcement​

How Manetu PolicyEngine controls access to resources.

Authorization Enrichment​

How per-server Rego policies translate low-level gateway signals into semantic operation and resource identifiers before the policy engine evaluates them.

Checkpointing​

How agent state is persisted and recovered.

Observability​

OpenTelemetry tracing, Prometheus metrics, and trace context propagation.

Runtime Tracing​

Syscall-level visibility into agent behavior via gVisor seccheck for security monitoring and debugging.

Payload Encryption​

AES-256-GCM encryption for Temporal workflow data at rest.

See also the Reference section for config-key-level detail behind several of these concepts, including TLS Trust Configuration.

Design Principles​

AgentVisor follows these design principles:

  1. Security by Default

    • Agents have no direct network access
    • All external access requires policy approval
    • Sandbox escape is prevented by container boundaries (and syscall interception in gVisor mode)
    • Real credentials never enter the sandbox (credential brokering)

  2. Transparent Compatibility

    • Standard HTTP libraries work unmodified
    • LangChain/LangGraph patterns work as-is
    • No special client code required

  3. Policy-Based Control

    • Declarative policy definitions
    • Fine-grained resource-level control
    • Audit logging for compliance

  4. Durability First

    • Automatic checkpointing
    • Recovery from failures
    • Human-in-the-loop support