Generative Agents (Stanford Smallville)

Stanford Smallville is the most landmark research in the A2A interaction field. Researchers created a virtual sandbox environment called Smallville where 25 LLM-powered AI agents live as "residents" autonomously. Each agent has unique personality, memories, and social relationships — they wake up, make breakfast, go to work, chat with neighbors, organize parties, and even spontaneously organized a Valentine's Day party.

The research introduced three key architectural components: (1) Memory Stream for storing experiences; (2) Reflection mechanism for extracting high-level insights from experiences; (3) Planning system for translating reflections into actions.

The paper won UIST 2023 Best Paper. Crowdsourced evaluation showed agent behavior was more believable than "humans pretending to be agents." Cited over 4,000 times, it pioneered the entire field of "LLM Agent Social Simulation."