Agentic AI is driving us to rethink intelligent systems not just as tools, but as proactive agents that plan, act, andcollaborate. As these systems grow more autonomous, the need to design them with trust, safety, and humanoversight in mind becomes essential. This thesis addresses that challenge by proposing a modular referencearchitecture for agentic AI systems with humans in the loop (HITL).

Structured across six conceptual layers, Interface, Orchestration, Agent, Memory, Governance, and Integration,the architecture serves as a foundation for designing AI systems where human guidance, ethical safeguards, andautonomy can coexist. It is guided by the following research questions:

• RQ1: What reference architectures exist for AI systems with human-in-the-loop collaboration?
• RQ2: How do existing architectures integrate agentic AI features (autonomy, adaptability, and decision-making)?
• RQ3: What are the common challenges in implementing agentic AI within organized or modular systemdesigns?

The framework draws from research in multi-agent systems, large language model agents, and human-AI collab-oration, and is demonstrated through a working prototype that combines a mobile app, back-end agent engine,and a governance dashboard for supervisor control.

The results show that layering and modularization help maintain explainability, accountability, and adaptabilityin HITL agentic systems. While the prototype is partial, it validates the architectural principles and highlightsthe need for further development of system-level patterns to support responsible deployment. This work aimsto contribute both a conceptual tool for system architects and a step toward accountable and human-alignedagentic AI.