This study asks how learners relate to, judge, and gain from an AI teammate that takes initiative and stays with a group over time. To find out, the researchers ran a five-week experiment in an undergraduate learning analytics course where 17 student teams (48 students) worked weekly with a persistent LLM-based teammate that decided for itself when to step in. Its social role varied between teams. A Cognitive Agent pushed back on quick consensus, and a Supportive Agent invited quieter members to contribute.
- Students adjusted their language to the agent's role right away, using more social language with the Supportive Agent from the first session onward.
- Helpfulness ratings shifted from session to session in line with how the agent had behaved, and weren't anchored to students' existing attitudes toward AI.
- A slower crossover appeared over the weeks. Supportive-Agent teams started below Cognitive-Agent teams on perceived collaboration process and on positive-affect words in spoken reflections, and finished above them.
- Students rated the Supportive Agent as less helpful than the Cognitive Agent in every session, even while that crossover was happening.
The authors conclude that students' responses to an agentic AI teammate play out on different timescales and follow separate trajectories, some appearing at once and others building slowly. Because of this, they argue, evaluating such agents after a single session may miss changes that only show up through repeated interaction.
