oasis

OASIS is an open-source Python framework for simulating social media platforms at scale, supporting up to one million AI-powered agents. Built by the CAMEL-AI research team and accompanied by an academic paper, it is a research tool for studying how information spreads, how polarization forms, and how crowds behave online, without running experiments on real platforms with real people.

The system works by creating a simulated version of platforms like Twitter or Reddit. Each agent in the simulation represents a user, driven by a large language model. These agents can browse posts, create content, follow other users, comment, like, dislike, search, and more. The simulation includes recommendation algorithms similar to what real platforms use, so agents encounter content based on their interests or on what is trending.

Researchers set up a simulation by defining profiles for their agents, choosing which actions agents are allowed to take, and running the environment for a set number of time steps. Each step, agents decide what to do based on what they see in their feed. Results are stored in a local database for later analysis and visualization.

The framework is designed for questions like: how does a piece of misinformation spread through a large network, and what slows it down? What conditions lead to echo chambers? These are hard to study on real platforms because researchers cannot control the environment or see everything happening.

Installation is via pip. An OpenAI-compatible API key is required to power the agents. The README includes a token consumption table to help estimate costs: running 100 agents for one time step uses several hundred thousand tokens, so expenses scale with simulation size and number of steps.