joonspk-research/generative_agents

Generative Agents is the research code that accompanies the academic paper "Generative Agents: Interactive Simulacra of Human Behavior." It is a working simulation of small computational characters, the project calls them generative agents, that try to act in believable, human-like ways inside a tiny virtual town named Smallville. You can watch them move around the map, go about routines, and interact with each other. The point is to study how language-model-driven agents can produce plausible social behavior over long stretches of simulated time. Under the hood the project is split into two pieces that run side by side. One is a frontend environment server, built as a Django web project, which renders the town map in your browser. The other is a backend simulation server, a Python program called reverie.py, which actually drives the agents' thinking and decisions by calling out to OpenAI's API. You start both servers, pick a starting scenario (for example, a three-agent setup with Isabella Rodriguez, Maria Lopez, and Klaus Mueller), tell it how many in-game steps to run, and then watch the agents act. Each step represents about ten seconds of in-game time. Simulations can be saved, resumed from a checkpoint, replayed step by step in the browser, or compressed into a tidier demo version with proper character sprites. This is a research artifact rather than a polished product. It needs an OpenAI API key, was tested on Python 3.9.12, and the authors warn that runs can get costly when many agents are involved. It is best suited for researchers, students, or builders who want to study or experiment with simulated AI-driven social behavior.