Automatic agenda extraction from minutes is useful for quickly grasping the contents discussed in the meeting. Extracted agendas can also be used as indexes for information retrieval. We previously proposed an extractive approach for this task, but our analysis revealed that a considerable number of human-created agenda expression did not appear as it is in the minutes. Therefore, in this paper, we propose an abstractive approach for this task. For each pair of question and answer segments, we apply the transformer-based sequence-to-sequence model to generate an agenda expression. We also incorporate a copy mechanism to the model. Our experimental evaluation showed that the proposed approach created more agendas similar to the human-created ones than extractive baseline in terms of ROUGE. We also conducted human evaluation and found that the proposed abstractive approach outperformed our previous extractive approach.