Young adults (18-30) frequently experience rumination after relationship conflicts—replaying conversations and emotions without resolution. While AI-enabled emotional support tools have emerged, most provide either validation or problem-solving approaches, rarely integrating both to support reflective processing. This thesis explores how multi-agent AI dialogue can help young adults break post-conflict rumination patterns by balancing immediate emotional validation with guided perspective-taking.

Through a Research through Design approach, I developed Bunny, a multi-agent AI system featuring two conversational agents—Protective Bunny for emotional validation and Objective Bunny for alternative perspectives—alongside a pattern reflection feature that synthesizes insights from users' conversation history. The design evolved through six prototype iterations informed by diary studies with five participants, culminating in an evaluation study with four participants using conversation analysis, semi-structured interviews, and User Experience Questionnaire (UEQ).
Findings reveal that participants valued the multi-agent dialogue structure and found pattern reflections insightful in identifying emotional triggers. However, the evaluation uncovered critical design tensions: the balance between exploratory questioning and users' need for concrete guidance, the challenge of appropriate conversational pacing, and the importance of giving users control over conversation endpoints. Participants showed strong interest in using Bunny for future conflicts (77.5% average on 0-100 scale, n=4), while expressing needs for more adaptive interaction strategies.

This work demonstrates both the potential and challenges of multi-agent conversational AI for relationship emotional support. The primary contribution is a design framework that explores how AI role configurations and interaction patterns can support post-conflict emotional processing. Through iterative prototyping and evaluation, this research identifies design tensions: the balance between validation and perspective-shifting, the challenge of appropriate conversational pacing, and the need for user control over therapeutic dialogue structure. These findings provide design implications for future AI systems supporting intimate emotional processing, particularly regarding the need for adaptive mechanisms that respond to individual users' emotional readiness and communication preferences.