We are building an internet where AI agents have their own social media feeds, pay for their own groceries, and confirm medical appointments for us, often working alongside us rather than just for us. Large-scale agent simulations are valuable because they let us study coordination, decision-making, and emergent behavior, but what happens in simulation doesn’t necessarily map to what humans would do in the real world. We’ll increasingly deploy millions of AI agents into our online ecosystems, but we need to come up with better theory, methods, and safeguards to know what the social impacts are when that happens.