Summary of Tier 4 — Autonomous Agent (Agentic Tier Framework v2026.1)
Tier 4 is the highest agentic tier, where an AI system plans and executes multi-step work autonomously against a defined objective. Human control is applied at the objective and materiality threshold, not at each step.
Defining Characteristics
- Autonomous planning of intermediate steps
- The AI chooses its own sequence of actions and tools to achieve an objective.
- Governance focuses on what goal is set and when to escalate, not how each step is performed.
- Cross-workflow action authority
- The agent can act across multiple systems, workflows, and decision types, rather than being confined to a single bounded process.
- No per-step human gating as the default
- Human review is triggered by materiality-based escalation rules, not by every action.
- Materiality calibration as a core control
- The definition of what is “material” (and thus must escalate) is explicit, versioned, and reviewed.
- Mis-calibration is the primary Tier 4 failure mode because it directly governs what the agent can do without human intervention.
Tier 4 is categorically different from Tier 3 because it plans: it decides what to do, in what order, and with which tools, rather than executing a predefined, human-authored step list.
Canonical Controls Required at Tier 4
A deployment is only defensible as Tier 4 if it implements all of the following controls:
- Materiality calibration as a first-class artefact
- Named, version-controlled, and reviewed on a defined cadence.
- Explicitly documents thresholds for escalation to human review.
- Reduced supervisory capacity monitoring
- Telemetry and monitoring that detect when human oversight is eroding or becoming ineffective.
- Operationalises the Reduced supervisory capacity risk class.
- Reversibility envelope
- For every action type the agent can take, there is a documented reversal plan and time window.
- Identifies when actions become hard or impossible to reverse.
- Real-time scope monitoring
- Detects when the agent operates outside its intended workflow surface, accesses unexpected data, or calls unexpected tools.
- Mandatory escalation paths
- Certain categories (e.g., regulator-facing, client-facing, irreversible transactions, materiality breaches) always require human gating.
- Tier 4 AI Council protocol
- AI Council approval is required at intake, for every material configuration change, and on a high-frequency cadence (typically monthly).