Skip to main content
Contribute

ARE concept

Progressive Autonomy for AI Agents

Autonomy should be earned through evidence, bounded by enforceable controls, continuously monitored, and reversible.

Definition

Progressive Autonomy is the principle that an AI agent earns greater operational authority as its measured reliability improves and loses authority when reliability deteriorates.

Autonomy is not a binary switch. It is a controlled spectrum of permissions, tools, data access, budgets, users, environments, and approval requirements. The objective is not maximum autonomy; it is safe, justified, and reversible autonomy.

Five stages

Observe

Proposed actions are logged and evaluated but not executed.

Draft

The agent prepares an action for a human to perform.

Act with approval

Execution requires explicit sign-off at the action boundary.

Act within scope

The agent executes autonomously inside a narrow, verified boundary.

Expanded authority

Scope widens only as reliability evidence accumulates.

Promotion and rollback

Evidence for promotion

  • Representative evaluation coverage
  • Stable success and policy-compliance rates
  • Known failure modes with bounded impact
  • Successful recovery and rollback tests
  • Complete action and authority telemetry

Triggers for contraction

  • Reliability or policy metrics cross a threshold
  • Drift or new failure modes appear
  • Observability becomes incomplete
  • A dependency or model changes materially
  • Incident severity or blast radius increases

Production example

A support agent begins in Observe mode, comparing its proposed responses with experienced agents. It moves to Draft after demonstrating stable recommendations, then to Act with approval for low-value refunds. Autonomous execution is allowed only for verified customer types and strict refund limits. If policy violations, customer complaints, or recovery failures exceed defined thresholds, the agent automatically returns to approval-required mode.

This approach prevents Reliability Debt by making authority expansion conditional on evidence rather than capability alone.