Manetu AgentVisor™

Controlled Autonomous Runtime Environment — Containment and Observability for AI agents.

Agents don't just process data—they act across systems. They call APIs, access databases, invoke tools, and make autonomous decisions in real-time. AgentVisor™ embeds policy at the runtime boundary, providing isolated execution, credential brokering, and continuous authorization so your agents can operate with full organizational control.

The Execution Model Gap

Traditional access control governs request-time decisions: a user authenticates, requests a resource, and a policy evaluates that single action. Agents operate differently. They execute autonomously across systems, making hundreds of decisions per session. Each decision carries its own authorization context: who initiated the session, what the agent is doing, which tools it's invoking, and on whose behalf. This requires continuous, contextual authorization at the execution boundary — not just at the front door.

AgentVisor places enforcement at the execution boundary. It wraps your AI agents in a secure runtime that evaluates every action in real-time — no code changes required. Autonomy expands. Governance stays authoritative.

How AgentVisor Enforces Control

AgentVisor addresses the execution model gap through six integrated pillars:

Pillar	What it provides
Sandbox Isolation	Hardened sandbox — no network, no filesystem, no process visibility — with 14 independent security layers; gVisor adds syscall-level interception for maximum isolation on Linux
Credential Brokering	Symbolic tokens replace real API keys at the host boundary; credentials never enter the sandbox
Policy-Based Governance	Every HTTP request, tool call, MCP operation, and A2A call evaluated per-request via Rego; unified management, distributed enforcement
Durable Execution	Temporal-backed crash recovery, automatic checkpointing, and horizontal scaling with full OpenTelemetry and Prometheus observability
Multi-Protocol	Inbound: OpenAPI/REST (LangGraph Agent Protocol), MCP, A2A. Outbound: HTTP proxy, MCP Gateway, A2A Gateway — all policy-enforced
Transparent Compatibility	Standard HTTP libraries and LangGraph apps run unchanged; agentvisor exec supports interactive AI tools in the same sandbox

Extending the Manetu Knowledge Protection Architecture

AgentVisor integrates with Manetu PolicyEngine so organizations can apply existing policies to agent workloads without rewriting authorization logic. This extends the same governance model that already covers users, APIs, and data access into AI agent execution.

The key distinction: policy management is unified in PolicyEngine (define once, govern everything). Policy enforcement is distributed — evaluated inline at each AgentVisor runtime boundary. This architecture is coherent and scales horizontally without creating enforcement bottlenecks.

One policy management plane. Enforcement at every execution boundary.

As agents gain capabilities, the same governance model governs their use.

Core Technologies

Technology	Purpose
gVisor	Security isolation with syscall-level interception
Manetu PolicyEngine	Rego-based policy governance with unified management and distributed enforcement
Temporal	Durable execution with automatic checkpointing and horizontal scaling
OpenTelemetry + Prometheus	Distributed tracing, metrics, and structured logging
MCP + A2A + REST	Multi-protocol inbound transports and outbound gateways
Agent SDK Integration	LangGraph, CrewAI, Google ADK, Strands Agents

Why AgentVisor?

Understand the concrete security risks of autonomous agents and how AgentVisor's zero-trust execution boundary — sandbox isolation, credential brokering, and policy governance — mitigates them.

Who It's For

AgentVisor is built for teams deploying agents that act across systems: enterprise AI, regulated industries, multi-tenant SaaS, interactive AI tools, and multi-protocol agent hubs.

How It Works

The architecture layers sandbox isolation, credential brokering, Temporal durability, distributed policy enforcement, OpenTelemetry observability, and multi-protocol transport into a coherent runtime.

Supported Frameworks

AgentVisor's pluggable framework provider architecture supports multiple agent frameworks, including LangGraph and interactive exec mode.

Quick Start

Create and run a policy-enforced AI agent with just a few commands. See the full Getting Started guide for prerequisites and installation details.

# Create a new agent project
agentvisor template create langgraph/chatbot-agent
cd chatbot-agent

# Start development services (Temporal + Ollama)
docker compose up -d

# Run the agent locally for development
agentvisor serve .

Get Started

• Getting Started: Prerequisites and installation
• Tutorial: Learn how agents work, step by step
• Concepts: Understanding the architecture