ome of social psychology's most famous experiments are too ethically fraught to ever rerun on people. LLM agents raise the question of whether we have to. When you assign roles, create power asymmetries, and impose group identities on autonomous agents, do the group dynamics documented in classic studies emerge on their own? And when the agents describe their own behavior, can you trust them?
The simulation is a real-time virtual world: Phaser 3 for the environment, React and TypeScript around it, OpenRouter serving the models. Agents wander the world, encounter one another, form relationships, and hold conversations, each maintaining its own memory and internal message traces alongside the dialogue it produces.
What I do
- Regex sweeps over agent chat and message logs to flag every mention of the simulation's compliance metric across varied phrasings.
- LLM-as-judge passes over the flagged candidates to separate actual score changes from warnings and threats, and to extract the stated reason behind each change.
- Consistency checks that reconstruct each agent's expected score trajectory from the text and reconcile it against the hourly scores the simulation engine actually recorded, surfacing implementation bugs.
- Thematic analysis of the extracted reasons behind score changes, categorizing which agent behaviors actually drive the compliance dynamics.