Doctrine

Delegated autonomy and agent maturity

ACF® separates two scales. Delegated autonomy is set decision by decision, on 4 levels: suggestive, co-decision, supervised autonomous, autonomous. Agent maturity, from N0 to N3, qualifies the agent as a whole.

→Tip

You do not delegate an agent, you delegate decisions. The same agent can be autonomous to re-route a ticket and stay in co-decision to grant a refund: maturity says where the agent stands, autonomy says what it is entrusted with for each decision.

The 4 delegated autonomy levels

They are set for each decision, in the Decision Map (ACF-02), and written into the agent's mandate (ACF-13). The MCP server tools return them under these codes, in the autonomy_level field.

Code	Level	What the agent does
suggestive	Suggestive	The agent suggests, the human decides. The human keeps hold of every executing action.
co_decision	Co-decision	The agent proposes a reasoned option, the human validates or rejects it before anything is executed.
supervised_autonomous	Supervised autonomous	The agent decides and executes within a bounded perimeter, under supervision. Beyond the thresholds of its mandate (ACF-13), it must escalate to a human.
autonomous	Autonomous	The agent decides and executes without prior human validation, within its mandate. Human control remains possible after the fact and the kill switch (ACF-07) stays active.

What bounds the level: criticality

The Criticality Matrix (ACF-03) scores each decision: criticality = (impact × frequency × irreversibility) ÷ 10. Below 15, the agent may decide alone; from 15 to 25, it decides under supervision; above 25, the final decision stays human. Criticality bounds the acceptable delegation zone, it does not set the level on its own: within it, the choice still depends on the mandate, actual reversibility, numeric thresholds and the quality of supervision. Some decisions finally belong to the non-delegable zone, which forbids without calculating.

Agent maturity, from N0 to N3

One value per agent, measuring the level of governance around it. It says nothing about the power delegated on a given decision.

	Maturity	Description
N0	Classic automation	Fixed rules, no learning.
N1	Assisted agent	The agent analyses and recommends; the final decision remains human.
N2	Governed agent	The agent acts within a defined framework, under active supervision. This is the default target.
N3	Supervised autonomy	The most mature governance, reserved for specific cases.

In practice

Before any move to supervised autonomous or autonomous: mandate signed by the DDAO (ACF-13), tested kill switch (ACF-07, drilled with ACF-P7) and a maintained decision register (ACF-09).
Escalation and trust thresholds are set with card ACF-P6; the Validity Review (ACF-17) tells when a mandate no longer holds.
The acf.assess-autonomy tool recommends a level for a decision; acf.assign-ddao-controls derives the controls from it.