We introduce an agent-based model of climate opinion dynamics powered by large language models. Each of our 212 computational agents simulates a specific member of the Pew Research Center’s American Trends Panel, initialized with synthetic first-person narratives constructed from 266 survey items across 17 waves. Building on Park et al.’s (2023, 2024) generative agent architecture, our LLM-based agents maintain memory streams, generate reflections from experiences, and plan discourse strategies. Agents participate in simulated social media discussions where they encounter climate information and misinformation. They deliberate about whether to engage, reply to one another, and update their beliefs dynamically as conversations unfold. Agents process natural language stimuli and generate human-like discourse, while the agent architecture is designed to help ensure behavioral coherence over time. Because both internal cognition and external behavior are fully observable, we can trace belief trajectories, identify persuasion mechanisms, and map information diffusion under full experimental control. We use this infrastructure to test motivated reasoning and selective exposure theories by manipulating message sources, scientific frames, and identity cues.