Control Domains¶
Eleven domains cover what agents do to data, to tools, to each other, and to the truth. Each domain scales by tier: the more autonomy an agent has, the more of the domain applies. Everything here inherits from Objective Intent, so read that first.
-
Prompt, Goal & Epistemic Integrity
Keeping instructions, goals, and "facts" trustworthy as they pass between agents.
-
Identity & Access
Per-agent identity and least-privilege permissions so authority cannot leak through hand-offs.
-
Data Protection
What data each agent may read, write, or move, and how it is protected in transit between them.
-
Document Extraction Integrity
Treating extracted document content as data to be checked, never as instructions to be obeyed.
-
Execution Control
Constraining what actions an agent can actually commit, and gating the irreversible ones.
-
Observability
Seeing what the fleet did: the flight recorder for multi-agent behaviour.
-
Supply Chain
Trust in the models, tools, and packages the agents depend on.
-
Privileged Agent Governance
Extra scrutiny for the agents that hold the keys.
-
Model Cognition Assurance
Assurance that the models still reason the way you validated them to.
-
Agentic Task Mandate
The scope of work an agent is mandated to perform, and the boundary it must not cross.
-
Emergent Risk Register
The living record of risks that only appear once agents interact at scale.
At single-agent scale these collapse into the ASO Core Controls: the same protections, one boundary.