OASIS (Open Agent Social Interaction Simulations)

OASIS imitates a social media platform and fills it with simulated users, each driven by a large language model (LLM).¹ It is built for studying how things spread online: information, group polarization, where opinions in a group drift to extremes, and herd behaviour.¹ It comes from the open-source CAMEL-AI project, and its 2024 paper has 23 authors, Ziyi Yang first and Jing Shao last.^1,2

The simulated platform has the parts that shape real ones.^1,2 Users have profiles and a follow graph that changes as they act.² Users can take 23 actions, including posting, commenting, reposting, quoting, following, muting and reporting a post.¹ Each user's feed is chosen by a recommendation system, either by interest or by a "hot score" of popularity.^1,2 Versions modelled on X and on Reddit are included.^1,2 The authors report that it reproduces information spreading, polarization and herd effects on both, and that larger simulated crowds produced stronger group dynamics and more varied opinions.²

You install it with pip as camel-oasis, supply a file of user profiles and a model, and step the simulation forward in Python; results go to a database.¹ It works with OpenAI models and others through the CAMEL framework.¹ The README gives a cost reference: 100 users for one step used about 336,000 input tokens on Alibaba's Qwen-turbo model.¹ Cost grows with users, steps and the share of users active each step, so a million-user run is a large undertaking.¹ Since 2025 a researcher can also interview a simulated user mid-run.¹

The README lists follow-on research built on OASIS, on collusion, fraud and marketplaces,¹ and MiroFish uses it as its simulation layer.³ The repository was created in November 2024 and was still being updated in October 2026.⁴ Each study has to test for itself how closely the simulated users behave like people on a real platform. Licence Apache-2.0.¹