Assessment model
AI Agent Reliability Maturity Model
Five levels for assessing how systematically an organization has implemented the Reliability Surface across autonomous agents.
| Level | State | Characteristics | Priority |
|---|---|---|---|
| 1 | Unmanaged | Identity and authority are undefined; actions are effectively unbounded and weakly tracked. | Inventory agents, tools, data, permissions, and effects. |
| 2 | Observed | Actions are logged, but authority remains broad and governance is mostly manual. | Define scopes, contracts, ownership, and minimum evidence. |
| 3 | Governed | Explicit authority and policy enforcement exist; approval gates and evaluations are operational. | Test failure containment, recovery, and policy coverage. |
| 4 | Accountable | Every consequential action is attributable to identity and policy; incidents produce traceable root cause. | Connect reliability evidence to authority decisions. |
| 5 | Adaptive | Authority expands and contracts automatically based on measured reliability and risk. | Continuously validate automated promotion and rollback. |
How to assess maturity
Score each dimension of the Reliability Surface independently. Use the lowest critical dimension to determine operating constraints: a system with mature observability but unmanaged authority is not Level 4. Record the evidence behind each score, the owner of each gap, and the conditions required to advance.
Authority should follow evidence
The maturity model is not a badge. It is an operating decision. Higher levels justify broader autonomy only when the system can demonstrate reliable behavior, enforce boundaries, reconstruct actions, contain failure, and recover safely.