This dissertation consists of four studies that explore how high school students interact with and learn from two designs of a text-based conversational agent (chatbot) during small-group science discussions with peers and the agents. The agents utilize natural language processing to send prompts to promote students’ understanding of ecosystems concepts and collaboration. The two agent designs differ in appearances and linguistic styles to resemble a less knowledgeable peer and an expert. The first two studies explore how students interacted with the agents as social partners. In Study 1, I found that interactions with the less-knowledgeable-peer agent generally contained more sequences of questioning and transactive exchange, which provided opportunities for reflection and reasoning. In Study 2, I examined how student groups attempted to remedy interactions with the agents in case the agents failed to interpret students’ intent, similar to how students would repair conversations with human partners. Student groups more often reframed and explained their reasoning to the less-knowledgeable-peer than the expert agent. Furthermore, such interactions were positively correlated with higher counts of systems thinking statements, which indicated an enhanced understanding of interconnections between human and natural systems. Study 3 presents a case study to illustrate how interactions with the agents varied with group compositions: emergent, mixed, and expanding prior science knowledge. In Study 4, I built on studies 1-3 to examine how students learned from the agents, using interaction patterns with the agents as mediators. Results confirm the affordances of certain interaction dynamics, such as transactive exchange, to deepen systems understanding. Overall, the studies provide converging evidence on the utility of agent designs to support interactions that are enriching for learning. I discuss implications for designing conversational agents as social partners to promote collaboration and systems thinking, with applications to other learning contexts.