Opponents and Partners: How Humans Interpret and Engage with Agents in Competition and Collaboration

Human-agent interaction is a dynamic process of mutual interpretation and response. When artificial agents participate in social contexts, whether as opponents in competitive games or partners in collaborative tasks, humans continuously observe their behavior, form judgments about their capabilities and intentions, and adjust their own actions accordingly. This thesis investigates how variations in agent design shape these interpretive processes and the quality of interaction that emerges, examining two complementary social contexts: competitive games and collaborative discovery. The central question guiding this research is: How do agent embodiment design, strategic sophistication, and conversational scaffolding influence human engagement, perception, and behavior in competitive and collaborative human-agent interaction? Rather than treating interaction as a static state to be measured, I examine it as an unfolding process shaped by the characteristics agents present to users, their physical form, behavioral patterns, and communicative style. To investigate competitive dynamics, I implemented and extended the Chef's Hat experimental framework, a multiplayer card game that provides a controlled yet socially rich testbed for systematic investigation. Within this environment, I manipulated agent embodiment design, comparing anthropomorphic robots to minimal black-box forms, all physically present, and strategic sophistication, from random baseline to reinforcement learning policies. Critically, agent strategies remained static, allowing isolation of how humans interpret and respond to different agent presentations rather than adaptation. To make these strategic differences legible, I developed novel visualization and descriptive methods that characterize agent behavior beyond simple performance metrics, revealing distinct strategic profiles ranging from transparent heuristics to sophisticated but brittle planning to robust, generalizable policies. Complementing this competitive investigation, I extended the research to collaborative contexts through a conversational music discovery system developed during an international study period. Here, participants co-created playlists with a social agent that employed structured dialogue and recommendation strategies to balance relevance with exploration. Unlike competition, where humans must model and counter opponents, collaboration requires trust, preference articulation, and openness to guidance. Findings indicated that conversational scaffolding and recommendation diversity significantly influenced exploration behavior, satisfaction, and the perceived quality of co-created outcomes, demonstrating that dialogue structure and mediation strategies are critical factors in collaborative human-agent interaction. Together, these studies constitute a dual-lens investigation of human-agent interaction dynamics. The competitive context reveals how humans interpret agent form and capability under adversarial pressure, while the collaborative context reveals how conversational design shapes trust and exploration under partnership conditions. Rather than claiming universal principles, this work identifies context-specific mechanisms while exploring whether certain design considerations, such as behavioral legibility, appropriate capability presentation, and scaffolded engagement, emerge across domains. The methodological contributions include validated experimental frameworks, novel visualization and descriptive methods for characterizing agent strategies, and empirical evidence demonstrating how agent design choices shape human experience across fundamentally different social configurations. These findings advance understanding of how artificial agents can participate meaningfully in human social spaces, informing the design of future interactive technologies that move beyond performance optimization to consider the character, interpretability, and social richness of interaction.