OASIS: Simulating Social Media With One Million AI Agents

By Prahlad Menon Published 2026-05-24 5 min read

What happens when you put a million AI agents on a simulated Twitter? OASIS (Open Agent Social Interaction Simulations) from CAMEL-AI answers that question — and the results are genuinely interesting for anyone studying how information spreads, opinions polarize, and crowds behave online.

What OASIS Does

OASIS is a scalable social media simulator where LLM-powered agents interact on platforms modeled after Twitter and Reddit. Each agent has a profile, preferences, and the ability to perform 23 distinct actions — following, posting, commenting, liking, disliking, reposting, searching, muting, and more.

The key claim: it scales to one million agents, enabling studies at a scale comparable to real social media platforms. That’s not a toy sandbox — it’s large enough to observe emergent phenomena that only appear at scale.

Why This Matters

Social media research has a data problem. You can study real platforms, but you can’t run controlled experiments — you can’t replay a scenario with different variables, inject specific content, or isolate causes. OASIS gives researchers a lab environment where they can:

Track misinformation spread — seed false information and watch how it propagates through the network
Study group polarization — observe how agents with different profiles cluster and radicalize
Analyze herd behavior — see how recommendation algorithms amplify trends
Test content moderation strategies — experiment with interventions before deploying them on real platforms
Simulate platform design changes — what happens if you remove likes? Change the algorithm? Add friction to reposting?

This is computational social science at a scale that wasn’t possible before LLMs made it feasible to give each agent realistic, context-dependent behavior.

Technical Architecture

OASIS is built on the CAMEL multi-agent framework and uses a gym-style environment API:

import oasis

# Create agent graph from profiles
agent_graph = await generate_reddit_agent_graph(
    profile_path="./data/reddit/user_data_36.json",
    model=openai_model,
    available_actions=available_actions,
)

# Make the environment
env = oasis.make(
    agent_graph=agent_graph,
    platform=oasis.DefaultPlatformType.REDDIT,
    database_path="./data/simulation.db",
)

# Step through the simulation
await env.reset()
await env.step(actions)

Key components:

Agent profiles — JSON-defined personas with interests, demographics, and behavioral tendencies
Platform simulation — Twitter and Reddit environments with authentic mechanics
Recommendation systems — interest-based and hot-score algorithms that determine what agents see
Action space — 23 actions covering the full range of social media interactions
SQLite storage — all interactions logged for post-simulation analysis
LLM backbone — uses GPT-4o-mini by default via CAMEL’s model abstraction (supports other providers)

The 23 Actions

Agents aren’t limited to posting and liking. The full action space includes:

Creating posts, commenting, liking/disliking posts, liking/disliking comments, reposting, following, muting, searching posts, searching users, viewing trends, refreshing feeds, and doing nothing — that last one matters because real users don’t engage with every piece of content they see.

This granularity is what separates OASIS from simpler agent simulations. The “do nothing” action alone introduces realistic engagement rates rather than forcing 100% participation.

Practical Use Cases

For researchers: Run reproducible experiments on social dynamics. Seed a conspiracy theory with 10 agents and measure how many of the remaining 999,990 encounter it within 100 time steps. Change the recommendation algorithm and re-run. Publish with reproducible results.

For platform designers: Test moderation policies, algorithm changes, and UI modifications in simulation before A/B testing with real users. What’s the polarization impact of removing public like counts?

For AI safety: Study how AI-generated content interacts with AI consumers at scale. When both the content creators and the audience are LLM-powered, what feedback loops emerge?

For educators: Demonstrate social media dynamics in a classroom setting. Students can design experiments, run simulations, and analyze results without needing access to real platform data.

How It Compares

OASIS occupies a unique niche. Most multi-agent frameworks we’ve covered are about task completion — agents collaborating to write code, do research, or manage workflows. OASIS is about social simulation: agents don’t have goals to accomplish, they have personalities to express.

The closest comparison is GenSim (Stanford’s “Smallville” paper), which simulated 25 agents in a virtual town. OASIS scales that concept by 40,000× and specializes it for social media dynamics.

Other multi-agent coordination frameworks like CAMEL (OASIS’s parent project), CrewAI, and PraisonAI are task-oriented. OpenSpace explored collective intelligence but at much smaller scale. OASIS is the first open-source framework purpose-built for million-agent social simulation.

Token Costs and Scaling

Running a million LLM agents isn’t free. At GPT-4o-mini pricing, even a single simulation step across all agents adds up. The practical approach is to run smaller simulations (hundreds to thousands of agents) for most experiments and reserve million-agent runs for validating scale-dependent phenomena.

The pip install camel-oasis package handles the infrastructure, but you’ll want to budget API costs carefully for large runs. Local models via Ollama could reduce costs dramatically for experiments where GPT-4o-mini quality isn’t necessary.

Getting Started

pip install camel-oasis
export OPENAI_API_KEY=your-key

Download a sample profile dataset from the repo and run the quickstart example. The documentation covers custom agent profiles, platform configuration, and analysis tools.

The Bigger Picture

We’re at an inflection point where AI agents aren’t just tools — they’re subjects of study. OASIS treats LLM agents as synthetic populations for social science research, which is a fundamentally different use case from everything else in the agent ecosystem.

The implications extend beyond academia. As AI-generated content floods real social platforms, understanding how synthetic agents interact at scale becomes a practical necessity, not just an intellectual exercise. OASIS gives us a sandbox to study that future before it fully arrives.