Tithi Mondal

Article8 min readUpdated

OASIS Social Simulation Explained for Founders

OASIS social simulation runs AI agents on a pretend social network. Learn how it works, what it needs, and what a small team can take from it.

By Tithi Mondal, UI/UX Designer

Read inहिन्दीবাংলাमराठीதமிழ்తెలుగుગુજરાતી

OASIS Social Simulation Explained for Founders: cover illustration

Imagine a coaching centre with 300 students that announces a fee change in its parent WhatsApp groups. Most parents read the message and say nothing. Then one parent writes an angry reply, 4 others agree, and the mood of the whole group changes. The fee did not change in those 10 minutes. The conversation did.

Researchers study this kind of group effect with simulators. OASIS is one of them. In one plain sentence: OASIS is an open-source social media simulator in which AI agents, driven by language models, behave like users on a Twitter-like or Reddit-like platform.

Checked in November 2026. Tools change quickly, so confirm details on the project's own page.

What is OASIS social simulation in plain words?

OASIS social simulation means running a pretend social network where every account is an AI agent. OASIS stands for Open Agent Social Interaction Simulations. The project comes from the CAMEL-AI open-source community.

Each agent has a profile, such as a name, a short bio and a description of personality. Each agent sees a feed of posts. Then a language model decides what the agent does next. The agent may post, comment, like, follow someone, or do nothing. All agents share one platform, so each action changes what other agents see later.

The project describes itself as "a scalable, open-source social media simulator". The research paper behind OASIS says the simulator can model up to one million users. That scale is a research claim about what the system supports. A small team would run something far smaller.

What problem does CAMEL-AI OASIS address?

CAMEL-AI OASIS addresses a research problem: how to study the behaviour of large online crowds without experimenting on real people. According to the project's documentation, OASIS is meant for studying information spread, group polarisation and herd behaviour.

Each term has a simple meaning.

  • Information spread is how a post travels from one person to many.
  • Group polarisation is when a group moves to a more extreme opinion after discussion.
  • Herd behaviour is when people copy the visible reaction of others, such as liking a post because many people already liked it.

Older simulators used fixed rules for each simulated person. The OASIS paper explains that such rule-based models are built for one scenario and are costly to adapt. A language model can respond to any post in ordinary language. So the same simulator can be reused for many scenarios.

The documentation lists 4 use cases: research simulations, interactive environments, content creation and behaviour prediction. Product validation for startups is not on that list. A founder who uses OASIS for a launch message is borrowing a research tool.

How does an agent based social media simulation work?

An agent based social media simulation works by repeating one loop: show each agent a feed, let the agent act, then update the platform. The OASIS documentation describes 5 core parts.

PartWhat it does
PlatformA Twitter-like or Reddit-like environment that holds accounts, posts and relationships
AgentsSimulated users, each with a profile and decisions made by a language model
ActionsThe things an agent can do, such as posting, commenting, liking and following
Recommendation systemThe rules that decide which posts appear in each agent's feed
Simulation engineThe part that moves time forward and activates agents

A run follows these steps.

  1. Prepare the profiles. The documentation says profile data is the first step. For the Reddit-style platform, a profile includes a name, a username, a bio, a persona description, age, gender, country and a personality type.
  2. Choose the platform and the allowed actions. The documentation lists many action types. One of them is "do nothing", which lets an agent pass the turn.
  3. Add the starting content. A developer can make a chosen agent publish a specific post. The documentation calls this a manual action.
  4. Run the steps. In each step the feeds refresh, the active agents read and choose actions, and the platform updates.
  5. Read the record. The documentation says all actions are logged in a SQLite database. SQLite is a small database stored as a single file.

OASIS also documents an interview action. A researcher can put a direct question to a single agent during a run, and the answer is saved in the database. The documentation says this action is designed to be used manually by researchers and developers.

What does OASIS need to run?

OASIS needs a developer, Python and access to a language model. OASIS is a code library, not an app with screens. The README shows installation as a Python package and a short example script.

  • A model. The quick start uses an OpenAI API key. An API key is a password that lets software use a paid service. The documentation says OASIS supports the models that the CAMEL framework supports, including models you host yourself.
  • Tool-calling support. The documentation notes that only models with tool-calling support can perform actions in OASIS. Tool calling is the ability of a model to choose an action from a defined list.
  • Profile files. You supply the agent profiles as files in the documented format.
  • A budget. Every agent decision is a model request. The README publishes a token usage reference for a 100-agent run. Tokens are the units of text that model services charge for. Cost grows with the number of agents and the number of steps.

The README states that the source code is licensed under Apache 2.0.

What can a non-research team realistically take from OASIS?

A non-research team can take 5 working ideas from OASIS, and most of them need no installation. The ideas matter more than the software for a small business.

  1. Reaction is a chain, not a sum. People react to your message and also to other people's reactions. Plan your launch post and your first 3 replies together.
  2. The feed decides who sees you. OASIS includes a recommendation system because reach is never equal. Ask where your message will really appear: a WhatsApp group, an Instagram feed, or a search result.
  3. Silence is an action. OASIS gives agents a "do nothing" option. Most real people also scroll past. A plan that assumes every reader responds is a weak plan.
  4. Profiles drive the result. OASIS asks for age, country, bio and persona for each agent. Write 8 short profiles of your real audience by hand. My guide on what a user persona is shows a simple format.
  5. Ask after exposure. The interview action questions an agent after the agent has seen the content. Do the same with real people: show the message first, then ask.

Here is a realistic use for this week. Suppose the coaching centre has a developer friend. The developer installs OASIS and prepares a small set of parent profiles. The developer then publishes the fee announcement as a manual post on the Twitter-like platform and runs a few steps. Finally, the developer uses the interview action to ask several agents what was unclear. The team gets a list of likely objections before the real message goes out.

Without a developer, do the paper version. Write the 8 profiles, write the announcement, and predict each person's first reply. Then show the announcement to 5 real parents and compare.

Where does simulating users with language models mislead?

Simulating users with language models misleads when a team treats the output as proof. Simulated reactions are generated by language models. They are a quick way to find questions and weak spots. They are not evidence of demand.

  • Made-up answers. A language model can write a confident statement that is not true. An agent may describe an experience that no real person had.
  • Bias. Language models learn from text written online. I would expect weaker results for groups that write less online, such as first-time internet users or people who prefer Bengali or Hindi.
  • Profile quality. The simulation is only as realistic as the profiles. Invented profiles give invented crowds.
  • A simplified platform. OASIS models Twitter-like and Reddit-like platforms. A parent WhatsApp group or a local Facebook group behaves differently.
  • Cost. Large runs consume many tokens. Set a spending limit with the model provider before you start.
  • Privacy. With a hosted model, your profiles and posts are sent to that provider. Do not use real customer names or phone numbers.

The OASIS paper reports that the simulator reproduced known social patterns in the authors' experiments. That is a result about group patterns in a research setting. That result does not say a simulation will predict the reaction to your product.

What stays with real users and the designer?

Real users still have to confirm the problem, and a designer still has to make and own the design decisions. No simulator takes responsibility for an outcome.

After a simulation or a paper exercise, I would do 3 things by hand. First, speak with 5 real people from the audience, using the method in my guide on user research for startups. Second, turn each confirmed objection into a design decision. If parents ask "what do I get for the higher fee", the answer belongs in the first screen of the message, not in a PDF attachment. Third, test the result with real people again, which my guide on usability testing on a small budget explains.

On edtech products I have designed, such as an app with live classes and progress reports, parents and students needed different information from the same data. A simulation can suggest that difference. Only conversations with parents confirm it. The same care applies to wording, which I cover in my post on UX writing and microcopy for apps.

If you have tested an idea and want help turning the findings into clear screens, you can read about my UI/UX design service.

Sources

Frequently asked questions

Who made OASIS?

OASIS comes from the CAMEL-AI open-source community. The research paper behind it is titled "OASIS: Open Agent Social Interaction Simulations with One Million Agents".

Is OASIS a ready-made app for marketers?

No. OASIS is a Python code library for simulations, and it needs a developer, agent profile files and access to a language model.

Which social platforms does OASIS simulate?

The documentation describes Twitter-like and Reddit-like platforms and says a customised platform can also be supplied. It does not simulate WhatsApp groups out of the box.

Can OASIS tell me if my product will go viral?

No. OASIS generates simulated behaviour from language models and profiles you supply. Use it to find likely objections, then test the message with real people.

Working on something like this?

Tell Tithi what you are planning. You get a clear scope, a fixed price and a timeline before any work begins. I reply to every enquiry within 24 hours.

Your details go only to Tithi and are used only to reply to you. No newsletters, no sharing.

Have a product in mind? Let us design it properly.

Tell me what you are building. I will reply with questions, a rough timeline and a quote.