Title: Compositional Governance: A Framework for Delegation and Scope in Agentic AI

Abstract:

As artificial intelligence transitions from static models to autonomous agents capable of initiating actions, collaborating, and assigning tasks, the conventional limits of software systems become increasingly indistinct. Conventional authorization and delegation models, which rely on fixed principals, explicit requests, and static scopes, are inadequate for regulating these agentic environments. Effective governance for Agentic AI requires more nuanced authorization semantics, including the ability for agents to inherit and pass on permissions, operate under time-bound authority, and synchronize via shared protocols. Current Identity and Access Management (IAM) infrastructures are unable to fully encapsulate the concept of agency, as they lack essential mechanisms such as recursive delegation, contextual boundaries, and dynamic scoping to serve as executable governance primitives. In contrast to established access delegation standards like OAuth 2.0, this approach treats delegation as a contractual obligation rather than a simple static token-based consent credential.

This study introduces a compositional governance framework featuring primitives essential for Agentic AI. It delineates various forms of delegation alongside their associated permissions and accountability implications, while also proposing a method for resource scope attenuation to limit the access envelopes of agents. These concepts are formulated as general relational definitions, allowing them to be integrated into existing authorization domains, such as financial systems. To facilitate this integration, we define a compositional operator that superimposes new agentic semantics—such as chains of recursive delegation—onto existing relational policies without requiring them to be rewritten. The validity of this framework is demonstrated through formal proofs and empirical assessments, establishing it as a rigorous yet practical foundation for ensuring accountable authorization within Agentic AI systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC