Introspection is central to identity construction and future planning, yet most digital tools for introspection approach the self as a unified entity. In contrast, Dialogical Self Theory (DST) views the self as a dynamic multiplicity of interacting "I-positions," distinct perspectives of the self that engage in inner dialogue. To bring this perspective into practice, we designed InnerPond, a multi-agent system that externalizes one’s I-positions as independent LLM-based agents in dialogue. Its design emerged through an iterative, metaphor-driven exploration that progressed from conflict-centered representations to spatial mapping of I-positions and ultimately to an integrative pond environment for introspection. In a user study with 17 young adults navigating career choices, InnerPond supported participants in exploring plural voices of the self, uncovering insights, mapping relationships among viewpoints, and developing metacognitive perspectives. Overall, this work suggests design implications for AI-supported introspection by shifting it from intrapersonal to inter-self dialogue.