Authority Before Execution: A Structural Control Invariant for Governable Autonomous Systems

As artificial intelligence systems cross from advisory roles into autonomous execution, failures previously treated as edge cases are becoming structural. These failures are commonly attributed to alignment gaps, insufficient oversight, or regulatory lag. This paper argues that such explanations misdiagnose the problem. The dominant failure mode in autonomous and agentic systems arises from a deeper architectural error: execution is routinely permitted before authority is explicitly resolved at runtime. We formalize authority before execution as a non-relaxable control invariant for governable autonomy. We show that widely adopted mechanisms—human-in-the-loop supervision, auditability, observability, compliance frameworks, and alignment techniques—cannot prevent cascading failure once execution has begun. Drawing from observed patterns across single-agent tools, multi-agent systems, enterprise automation, research platforms, and frontier models, we demonstrate that autonomy scales faster than accountability under execution-first architectures. We conclude that some autonomous actions must be rendered structurally impossible, regardless of technical feasibility, if autonomy is to remain governable. Without this invariant, governance collapses into post hoc attribution, and responsibility is assigned only after irreversibility has occurred.