Engineering Reliable Agentic AI: Decision-Making in Real-World Deployments

Large language model (LLM) agents are increasingly deployed to give everyday guidance whose correctness depends on rules that users do not know and on personal facts that the agent must first elicit. At each turn, such an agent must choose whether to ask a clarifying question, act through a tool, answer, or decline. This paper presents Rikanverse®, a Proactive, Multilingual, Voice-Native Agentic AI Application, built on a reasoning-and-acting (ReAct) architecture and designed, built and deployed end-to-end solo by the Author. Rikanverse® serves both sides of the market: tenants describe their needs by voice or text and receive matched listings with practical guidance, while landlords register listings by speaking instead of completing forms. The agent treats clarification as an explicit action, acts through a small, typed tool set, maintains persistent session state, and checks every draft answer before release, declining when it cannot answer reliably. The paper describes the system, reports preliminary user evidence from a survey of 30 respondents, and proposes an evaluation protocol in which the trade-off between correct answers and declines serves as the central measure of reliability. The work shows how agent design choices about asking and abstaining can be made explicit and measurable in a deployed system.