<![CDATA[In the era of disruptive technology, increasing reader engagement rates has become a key factor for online media industries in Indonesia, as it directly impacts advertising revenue. To address this challenge, generative AI has emerged as a critical technology that can deliver more personalized, relevant, and adaptive reading experiences. This study introduces a prototype news recommendation system based on an AI Agent designed to provide adaptive and sustainable reading experiences. The system integrates several components, including short-term and long-term memory, association, similarity, and a generative mechanism powered by Large Language Models (LLMs) at the core of the agent. The system was evaluated using two prompting approaches: static prompting, which keeps the recommendation prompt fixed, and adaptive prompting, in which generative recommendations clicked by users dynamically update the prompt fed to the LLM. The evaluation was conducted across six questionnaire-based metrics from 49 respondents: diversity, novelty, serendipity, curiosity, filter bubble, and context coherence. Six open-weight LLMs from the Ollama platform were tested and categorized as large LLMs (>100B parameters) and small LLMs (<20B parameters). The experimental results indicate that adaptive prompting consistently improves contextual coherence and reader curiosity. Large LLMs achieved the highest scores across nearly all metrics, particularly serendipity and curiosity, demonstrating their potential to deliver adaptive and sustainable reading experiences that increase reader engagement. These results provide important contributions to the development of agentic AI in news recommendation systems, paving the way for more adaptive, contextual, and personalized reader interactions.]]>
