A Governance Framework That Splits Allowed Autonomy Levels From Autonomous Capability Levels

arXiv 2607.23438·medium signalXBlueskyLinkedInCopy link

This paper (arXiv 2607.23438) argues that autonomy debates conflate what an agent can technically do with what it should be permitted to do, and proposes separating Allowed Autonomy Levels (AAL) — authorized autonomy given risk, oversight, and accountability — from Autonomous Capability Levels (ACL), the agent's inherent technical ability. It lays out a level ladder spanning reactive execution, decision support, supervised action, goal-directed autonomy, and delegated operational authority, describing how control, reversibility, and accountability shift as you climb. The framework is demonstrated on a deployed enterprise data-engineering agent that scores high on capability but is deliberately pinned to a lower allowed autonomy based on reversibility and organizational readiness — a concrete pattern for teams writing agent permission policy.