In an era marked by growing reliance on interdisciplinary collaboration, online platforms have become crucial facilitators of knowledge integration. Yet, interdisciplinary teams consistently encounter persistent barriers — terminology mismatches, fragmented attention, and delayed trust-building — that impede effective communication and innovation. While conversational agents hold promise for enhancing human-AI collaboration, their application in complex, knowledge-intensive interdisciplinary settings remain underexplored, particularly in balancing users’ cognitive demands and engagement. This study investigates how conversational agents can optimize online collaboration through tailored interaction designs. For the purpose, semi-structured interviews with 20 students and a series of controlled experiments involving 32 participants from four disciplines were conducted. We evaluated four conversational agent configurations (i.e., approachable / professional visual embodiments; proactive / suggestive interaction modalities) based on eye-tracking experiments, questionnaires, and user observations. The results revealed that proactive conversational agent design can expedite users’ attention allocation (60.57% faster first-visit time to critical information regions; 33.14% fewer saccades), but increase cognitive processing difficulty (28.70% longer saccade latency), whereas suggestive designs reduce processing complexity at the cost of attention dispersion. Approachable visual had a negligible impact, as users prioritized textual content. Interaction modes significantly influenced collaborative efficiency (F = 6.08, p = 0.02), moderated by participants’ communication competence (F = 3.88, p = 0.05). These findings can provide actionable guidelines for AI-mediated collaboration, and facilitate the design of conversational agents with higher engagement and collaboration quality in interdisciplinary contexts.