Design and Implementation of Secure Agentic AI Systems Using Token Vault Architecture for Real-Time Enterprise Applications
Abstract
Agentic AI is moving beyond conversational assistance and is increasingly being used to carry out enterprise tasks such as invoking tools, working with repositories, scheduling activities, and coordinating actions across business systems. This capability is useful, but it also introduces a security concern when the same agent runtime has access to reusable credentials or long-lived access tokens. This paper presented a Token Vault-based architecture. The architecture separated credential handling from the reasoning and orchestration functions of the AI agent. Long-lived credentials remained outside the agent runtime. Short-lived tokens with limited scope were issued only after identity, policy, and approval checks. Human-in-the-loop authorization was also included. A separate execution layer interacted with enterprise APIs and external services. The layered architecture separated planning, context preparation, inference, execution, validation, and observability. The study evaluated the approach through controlled enterprise workflow scenarios. It compared the Token Vault path with direct-token execution. The comparison showed improved credential isolation, scoped authorization, security control, and auditability. Token mediation added a limited amount of processing. The study showed that reasoning and credential ownership were separate concerns. This separation provided an additional security boundary for enterprise agentic AI. Future work will test the approach with larger workloads, additional enterprise services, and multi-agent environments.