Even though many of our daily activities are sequential in nature, this sequencing has received only little attention in the dual-tasking literature. However, an important source of dual-task costs is missed here. It has been shown that implicit sequence learning suffers when a serial reaction time task (SRTT; Nissen & Bullemer, 1987) is combined with a secondary task. Thus, the amount of sequence learning could serve as a marker for the limitations of dual-tasking. Schmidtke and Heuer (1997) suggested that participants tend to integrate both task streams and that implicit sequence learning under dual-tasking conditions is hampered to the extent that they are uncorrelated – resulting in extraordinarily long integrated sequences which are hard, or even impossible, to learn. While this assumption implies that the transitions between across-task events are learned concurrently within- and across-trials, former results of Röttger, Haider, Zhao, and Gaschler (2019) hint at the primacy of learning the contingencies between the SRTT target and the secondary task stimulus within a trial. The goal of the present three experiments was to further investigate this observation. More specifically, the experiments test whether implicit sequence learning in dual-task situations is hampered to the extent that across-task predictions within a trial interfere with within-task predictions across trials (i.e., associative chaining within the SRTT). For this purpose, we manipulated the contingency between each visual element of an 8-element 2nd order SRTT and the two secondary task tone stimuli in different ways and assessed to what extent sequence learning occurred. It turned out that sequence learning was the more pronounced the easier it was to learn all within-trial SRTT-tone pairs first. It was strongly reduced if half of the SRTT-elements were randomly paired with the tones. This outcome suggests that one important source of dual-task costs is a tendency to represent both tasks within one single task set – even when they are uncorrelated – preventing any reduction of the within-trial prediction error.