This paper examines agentic AI systems that internally delegate tasks by instantiating sub-agents. It frames the resulting risks as problems of delegation, authority, and accountability rather than autonomy or emergent intelligence. The work maps these systems to the requirements of ISO/IEC 42001:2023 for Artificial Intelligence Management Systems (AIMS) and outlines governance controls applicable under existing standards, including bounded delegation, non-delegable authority, admissibility prior to execution, lifecycle revocation, and audit-sufficient evidence. The paper is intended to support enterprise deployment, regulatory scrutiny, and continual improvement without relying on speculative assumptions about future AI capabilities.
Anthony MacFarland (Mon,) studied this question.