A million LLM agents, one synthetic social network

What it does

OASIS is a social media simulator that turns large language models into platform users. Researchers can spawn agents on Twitter- or Reddit-like environments to observe emergent phenomena such as information cascades, group polarization, and herd behavior. It handles the full stack: agent profiles, social graphs, content generation, and platform-specific recommendation feeds.

The interesting bit

Instead of scripting crowd behavior, OASIS lets algorithmic feedback do the work. Agents choose from 23 actions—follow, mute, repost, search, create_post—while built-in interest-based and hot-score recommendation algorithms shape what they see next. The project also publishes surprisingly frank token-consumption and Yuan-cost tables for Qwen, which is a rare dose of financial realism in agent-framework READMEs.

Key highlights

• Claims architectural support for one million agents, with environments mimicking Twitter and Reddit dynamics.
• Agents draw from a palette of 23 actions, including social (follow, mute), content (create_post, create_comment), and discovery (search, trend, refresh).
• Includes recommendation algorithms (interest-based and hot-score-based) that drive content discovery rather than leaving it to chance.
• Provides token and cost benchmarks for 100 to 10,000-agent runs using Qwen models, making API burn predictable.
• Recent refactor adopts a PettingZoo-style interface and allows per-agent customization of models, tools, and system prompts.

Caveats

• The “one million agents” figure is a scalability target, but the README only supplies concrete token and cost data for up to 10,000 agents, leaving the price tag of full-scale runs unstated.
• All documented benchmarks assume external cloud LLM APIs (OpenAI, Qwen), meaning simulation costs scale directly with every agent action.
• The quickstart examples and current API surface are heavily centered on Reddit, making the maturity and feature parity of the Twitter/X platform mode less obvious.

Verdict

Social scientists, misinformation researchers, and recommendation-system engineers should bookmark this. Skip it if you are looking for cheap chatbot automation or a lightweight offline sandbox.