The Authority Problem in Agentic AI: Why Execution Requires External Admission Boundaries

This preprint introduces the Authority Problem in agentic AI: the problem of final authority over execution in systems that can initiate or mediate consequential actions. It argues that execution-capable AI systems require external admission boundaries: pre-execution control points where execution cannot proceed without an independent allow decision, denial is fail-closed, and an admission record exists before execution.The paper distinguishes external admission from guardrails, monitoring, audit logging, rollback, policy-as-code, and human approval loops. It also describes the Surrogate Boundary Test as a diagnostic method for determining whether a claimed execution boundary actually separates execution from final authority.This work does not claim comprehensive AI safety, legal compliance, complete security, or prevention of all harmful outcomes. It is a classification and control-model contribution for non-bypassable execution admission in agentic systems. License and use notice: This artifact is released under CC BY-NC-ND 4.0 It may be read, cited, linked, and referenced with attribution. No commercial use, modified versions, derivative specifications, operational claims, service claims, or implied endorsement are granted by this publication. No implementation, operational, commercial, licensing, evaluation, conformity, equivalence, endorsement, or authoritative interpretation rights are granted by this publication. Any such use requires separate written permission from the author. Licensing and interpretation framework:https://doi.org/10.5281/zenodo.18381146 Current clarification version:https://doi.org/10.5281/zenodo.19846720 Project page:https://ai-admissibility.com/