Social bots influence opinion polarization in online communities by mimicking human behavior, posing a significant threat to cognitive security. Traditional studies on social bots have often relied on rule-based or statistical simulations, which fail to capture the complexity and realism of social bot behavior. To address these limitations, we introduce BotSphere, a multi-agent simulation framework powered by large language models, to study how social bots manipulate public opinion in online networks. BotSphere integrates a timeline component, complex networks, and a multi-agent framework to simulate the dynamics of opinion manipulation. We develop LLM-based human-like agents to represent social media users and embed social bots to influence opinions. Through simulation experiments, we explore how the number of social bots and their narrative manipulation strategies affect public opinion. Our results demonstrate that BotSphere provides deeper insight into the role of social bots in shaping public opinion, offering a more realistic and nuanced understanding of social media dynamics compared to traditional models.