Doctrine

The 4 founding principles

ACF® rests on four axiomatic principles, independent of the technical protocol used underneath. They apply to any agentic system in production, regardless of the client (Claude, GPT, Gemini, proprietary model) or the protocol (MCP, ACP, Operator, etc.).

→Tip

Principles are not ranked. P3 without P1 produces an agent that decides on its own but is traceable — not enough for a regulated use. P1 without P3 produces an agent supervised in intent but undefensible in practice. The four hold together or not at all.

P1 — Decision / execution separation

The agent may execute; it cannot decide alone on strategic-critical actions (contractual commitment, financial gap above a threshold, international data transfer, irreversible action on a real asset). These decisions are always human, or suspended.

P2 — Non-delegable zones

Some decisions are never delegable, whatever the agent’s autonomy level and the organisation’s maturity. They are enumerated in each organisation’s Agentic Constitution (card ACF-03).

P3 — Traceability and interruptibility

Every agentic action is logged in a cryptographically signed register (card ACF-08). Every action can be interrupted at any time via an operational kill switch (card ACF-06) whose effectiveness is tested.

P4 — Living governance

The governance framework evolves with the agents’ capabilities. A formal review at least quarterly (cards ACF-05 and ACF-10) adjusts autonomy levels, escalation thresholds, and non-delegable zones.

Why axiomatic?

ACF® principles are posited as axioms — not derived from anything else — so an organisation can recognise or reject them as a whole. A debate on “should we trace decisions?” does not lead to a deployment. A decision on “do we adopt the ACF® framework?” unblocks the next step. This is a deliberate methodological posture.