With the advent of large language models (LLMs), the concept of artificial intelligence (AI) agents capable of decision making and action execution has gained prominence, particularly as LLMs demonstrate increasing proficiency in tool use and task planning. Following these advancements, the established methods used for task-oriented dialogue (TOD) systems have undergone a paradigm shift by integrating LLMs’ revolutionary language understanding and reasoning skills with enhanced instruction following and response generation abilities. At the intersection of these developments, we motivate the evolving field of conversational AI agents, LLM-based conversational systems that combine advanced language abilities with agentic qualities to engage in more dynamic, context-aware, and task-oriented interactions. This “Perspectives” column examines the evolving landscape of conversational AI agents and discusses recent advancements and future directions, especially focusing on conversational task completion, which requires function or application programming interface (API) calling. We argue that while LLMs have significantly enhanced these conversational agents, challenges in multiturn context management, controllability via policy following, personalization, user overreliance, and comprehensive evaluation methodologies remain critical areas for future research and development.