A Conceptual Framework for Authentication in Agentic AI Ecosystems: Protocol Analysis and Taxonomy

Abstract

Agentic AI platforms launch short-lived software agents that act with delegated authority across services and organizations. Existing authentication protocols were designed for traditional web applications and long-lived microservices, not ephemeral autonomous agents, creating uncertainty about their applicability to agentic scenarios. This paper systematically analyzes authentication mechanisms for agentic ecosystems, including SPIFFE/SPIRE for workload identity, OAuth 2.0 and OpenID Connect for user delegation, Token Exchange for delegation chains, DPoP for sender-constrained tokens, CIBA for human step-up, and decentralized identity approaches for cross-domain portability. We propose a five-pattern taxonomy organizing these mechanisms by interaction type: user-to-agent, orchestrator-to-agent, agent-to-internal service, agent-to-external API, and cross-domain federation. For each pattern, we evaluate capabilities and limitations based on specifications and implementations, identify integration prerequisites, and highlight operational challenges. To validate the feasibility of protocol composition, we implement a hybrid prototype that demonstrates agent-to-internal-service authentication. Measurements show that sender-constrained tokens add 52ms of cold-path and 11ms of warm-path overhead, with token caching providing a 4.6× speedup. Our analysis indicates that protocol compositions can address core agentic requirements (ephemerality, delegation, replay prevention, auditability) within a single trust domain, while identifying critical research gaps, including multi-hop delegation semantics, federation automation, and high-churn performance.