An Identity, Delegation, and Accountability-Based Action-Centric Identity Management Architecture for Agentic AI

As Agentic AI systems increasingly act autonomously on behalf of humans, conventional identity and authorization models designed for static users or services become insufficient. Existing approaches, such as IAM and OAuth, focus on access validity rather than on identifying who acted, under whose authority, and with what accountability, which is critical when autonomous agents independently plan and execute actions. This paper reframes identity management in Agentic AI as an action-centric problem structured around independent agent identity, verifiable delegation, and action-level accountability. From this perspective, we formalize nine core requirements (R1∼R9), including ephemeral identity life-cycle management, re-delegation control, machine-verifiable validation, and human-in-the-loop governance. To address these requirements, we propose a minimal identity management architecture that integrates persistent agent identity, verifiable delegation credentials, and cryptographic action binding. The architecture enables each autonomous action to be explicitly justified, validated, and auditable. We further evaluate representative Agentic AI protocols-MCP, ACP, A2A, and ANP-against the derived requirements to identify structural gaps. This work provides a reference architecture for accountable autonomy in Agentic AI.