Online education creates new opportunities for learners with different characteristics leading to sharply increasing enrolment in the last few years. However, the lack of learner engagement and the corresponding high attrition rates in online courses remain a major challenge in the field. Scaffolding-based conversational agents promise to address this problem by interacting with each learner individually, similar to educators' scaffolding behaviour. Conversational agents (CAs) are one of the biggest trends in digital transformation and one of the most common applications of AI. These are computer programs that can interact with humans in natural language. A growing body of evidence suggests that these programs can change a student's commitment to learning. Conversational agents with recent natural language processing (NLP) capabilities can offer dynamic scaffolds to learners helping them to gain new knowledge and being more engaged in the learning process. So far, there have been a few studies that prove CAs have a positive impact on learning success. However, it is hardly clear how much CAs have impacted learner engagement and satisfaction. Some studies have checked engagement with self-reported data, but none has used sensor data. From learner engagement theory, however, it is known that engagement is better measured with sensor data. The aim of this paper is to conduct a laboratory experiment to measure the impact of conversational agents on learner engagement in online learning. In a second step, qualitative data from a survey after the experiment will be used to determine reasons for student satisfaction or dissatisfaction and to find out how students perceived different CAs during online learning. To test the hypothesis, I built a dynamic scaffolding conversational agent named Sara and compared it with a static, one-size-fits-all scaffolding conversational agent. The CAs are included in two online video lectures and provide.