-
Notifications
You must be signed in to change notification settings - Fork 0
Four Layer Architecture
Agent Hypervisor is organized as four deterministic layers. The architecture is intended to provide ontological security: the agent is governed by what can exist in its world, not by a probabilistic runtime judgment about whether a tool call sounds safe.
Layer 0 is infrastructure isolation: containers, process boundaries, filesystem limits, network controls, and other physical constraints. It is the bedrock for actions that must be impossible at the system level.
Layer 1 defines the vocabulary of actions the agent may ever propose. Tool surfaces are specialized into capabilities, schemas, and parameter constraints. An action outside the base ontology is not merely denied; it is absent from the agent's world.
Layer 2 projects the base ontology into the current context. It accounts for actor, role, task, trust channel, environment, and triggering event. This layer determines which capabilities are visible or available right now.
Layer 3 is the deterministic runtime engine. It evaluates proposed actions
against compiled policy artifacts, provenance chains, taint state, and execution
mode. Outcomes include allow, deny, ask, and simulate.
The architecture mirrors classical virtualization, but it virtualizes meaning, actions, and consequences instead of CPU, memory, and IO. The agent runs inside a bounded semantic world.