Containment Plane

The Containment Plane is the set of controls that lets an institution keep an AI agent within bounds. It answers five questions about any agent in operation: can it be bounded, observed, interrupted, attributed and recovered from?

It is called a plane because it cuts across every layer of an agent system. Identity, permissions, orchestration and trust each contribute controls, and the plane is where they add up to one answer about failure. In the Agent Layer Framework, containment is what governs failure. Operational Alignment defines the outcomes an institution needs: containment, observability, interruption, attribution, recovery and learning. The Containment Plane is the control surface that delivers them.

Two shifts are reshaping it:

• Controls are moving outside the agent. The most consequential controls increasingly sit outside the agent's own reasoning and runtime, down to independent hardware that can stop an agent regardless of what the agent decides. The value of independent interruption rises as agents act in the physical world, where mistakes are harder to reverse.
• The perimeter is dissolving. Containment assumes the agent runs inside infrastructure one institution controls. When useful AI can execute across devices, networks, owners and jurisdictions that no single party controls, the question becomes where the perimeter is at all.

The Containment Plane is an Alignment-force concept: whoever can credibly contain agents earns the authorization to deploy them in high-stakes settings.

Related: The Agent Layer Framework, Signal Brief: Alignment Moves Into Silicon, Signal Brief: When the Containment Plane Loses Its Perimeter, Alignment Sovereignty and Authorization Capacity.