Building sophisticated, goal-oriented AI Agents at scale introduces profound architectural and reliability
challenges. This talk dissects Sendbird's journey in engineering a production-grade AI Agent platform
designed for complex, multi-step actions across communication channels.
We will cover the foundational Agentic Architecture while specifically focusing on the engineering required
to achieve predictable outcomes. This involves two critical areas: ensuring reliable function calling where
the LLM accurately translates user intent into structured API arguments for complex backend systems,
and establishing accurate instruction following for internal Actionbooks, text-based directives crucial for
service control and consistency.
Furthermore, we will examine how extending these capabilities to Voice AI Agents compounds the
challenges, demanding solutions for real-time latency and conversational flow management atop existing
reliability needs. We conclude with our production solutions for operational scale, including cost
optimization through caching and robust systems for hallucination mitigation and continuous performance
evaluation.
Attendees will gain practical, engineering-focused insights required to build reliable and predictable AI
Agents that confidently interact with core enterprise infrastructure.

Kibeom Lee, Engineering Manager (AI/ML), Sendbird