CoSAI AI SRF §3.2

Operating Models

The fastest way to unblock a deployment is to know which responsibilities you keep and which the provider carries. That split shifts significantly depending on whether you are running IaaS, AI-PaaS, Agent-PaaS, or AI-SaaS. Select a model to see the full layer-by-layer breakdown of who owns what.

Loading...

Companion explanation; not a fifth CoSAI operating model. SRF v1.0 still has four models. Autonomy classifies how independently an agent may act; it does not move the accountable persona. For Agent-PaaS and IaaS especially, ask where authorization is actually enforced, not only where the agent executes. Proposed for the next SRF paper →
Enforcement planes. North-south is agent to model provider (often L5 policy and spend). East-west is agent to agent (Agent-PaaS identity and delegation). Agent-to-tool is MCP or other tools (L3). On Agent-PaaS the provider typically owns the orchestration runtime that can halt a swarm; the customer owns agent definitions and which tools are granted. On IaaS the customer owns nearly the entire enforcement path. On AI-SaaS the provider owns most runtime enforcement; the customer still owns L1 outcomes. See enforcement plane.