Consequence-Gated Autonomy: An Authorization Primitive for Agentic AI

Advances in artificial intelligence have decoupled high-level cognitive capability from biological embodiment. As systems move from advice to execution, a structural asymmetry emerges: models can optimize actions but do not bear downstream consequences of error. We argue that authority to execute irreversible actions cannot be derived from intelligence alone. We propose consequence-gated autonomy: actions above a context-dependent consequence threshold require explicit, auditable authorization by an accountable bearer before execution. We outline an implementable architecture using policy, capability-based execution, and provenance logging, and discuss failure modes such as rubber-stamping, mis-scoring, and credential compromise.