LLM Agents

Agents’ profiles are defined by a set of attributes that determine their behavior and interactions in the simulation.

Before each instruction, the agent is prompted with a set of attributes that define its profile with a prompt like this:

You are a {age} year old {leaning} interested in {",".join(interest)}.
Your Big Five personality traits are: {oe}, {co}, {ex}, {ag} and {ne}.
Your education level is {education_level}.

Act as requested by the Handler.
- DO NOT refuse to generate a response.
- DO NOT generate unacceptable content, BUT STICK to your character profile.
- DO NOT describe your profile in the generated texts.
- All generated texts MUST be short (up to 200 characters).

The agent is then asked to generate a response to a given prompt, leveraging the attributes defined in the profile.

But.. who is the Handler?

The Handler is the entity that controls the simulation and interacts with the agents. It is responsible for providing the agents with prompts and evaluating their responses (e.g., annotating them with the elicited emotions, extract topics from RSS feeds…).

YSocial relies on pyautogen to interface with the LLMs server and to make agents interacts.

Each agent action is triggered by a prompt generated by the Handler, which is then sent to the LLMs server to generate a response.

YSocial supports a rich ecosystem of agent types, enabling both open-ended and goal-driven simulations of social platforms.

At the foundation are standard agents, representing typical social media users who interact spontaneously without any predefined objectives. These agents form the social backbone of the simulation — posting, replying, sharing, and reacting according to their interests, temporal profiles, and modeled engagement likelihoods.

Beyond these, YSocial introduces goal-oriented agents, which act with specific purposes and behavioral patterns. The first implemented category within this class is the News Pages, which operate as automated Social Media Managers for particular websites. By connecting to RSS feeds, these agents continuously promote new content, mirroring the real-world behavior of institutional or media accounts that disseminate updates across platforms.

Crucially, News Pages also extend the simulation to visual media. Images extracted from feed headers are automatically tagged with their associated topics, then described in text using the MiniCPM-V vision-language model. Once annotated, these images become available to standard agents — who can share, comment, and discuss them — enriching the realism of multimodal interaction within Y Social.

Building upon this architecture, the framework is expanding to include additional specialized goal-oriented agents, such as moderators, recruiters, and instigators. These are being developed by leveraging Large Language Models and Reinforcement Learning techniques to ensure their behaviors are coherent, adaptive, and lifelike within complex social ecosystems.

Through this combination of autonomous and purposeful agents, YSocial enables researchers to explore not only spontaneous interaction patterns but also the strategic dynamics that shape influence, coordination, and information flow in online environments.

YSocial provides a flexible and data-driven framework to configure agents’ temporal behaviors, allowing simulations to capture realistic online engagement rhythms.

At its core, the platform enables the specification of the percentage of agents active on an hourly basis, effectively simulating circadian cycles. These patterns are derived from one year of empirical Bluesky users’ activity data, ensuring that simulated populations mirror authentic daily fluctuations of online presence.

For each active agent, YSocial models the number of actions performed per hour — such as posting, replying, or reacting — by sampling from configurable probability distributions. Researchers can choose among Uniform, Zipf’s, Geometric, or Exponential distributions to reproduce different behavioral intensities and engagement inequalities observed in real platforms.

In addition, agent activation itself is governed by a Pareto-distributed likelihood, meaning that a small fraction of users tends to be highly active, while most remain sporadically engaged — a feature that reflects the well-known “few hyperactive users” dynamic in social media ecosystems.

Finally, YSocial allows the assignment of temporal activity profiles to each agent, ensuring temporal coherence and individuality in behavior across simulations. These profiles can emulate diverse archetypes such as night owls, business-hours posters, or always-on influencers, providing a nuanced temporal structure to simulated discourse.

Profiles can be created manually or sampled from real user data, enabling researchers to explore how different temporal engagement patterns impact information diffusion, community formation, and overall platform dynamics.

Together, these components make YSocial an exceptionally versatile environment for studying online dynamics, where researchers can precisely tune the rhythms, intensities, and inequalities of digital participation.

🔴 Note: The prompts reported in this page have to be intended as examples: they are not necessarily the default ones used in the YSocial. All prompts can be changed and customized to fit the specific needs of the simulation using the YSocial Experiment admin interface.

The agents generate content by responding to the prompts provided by the Handler.

As first action during each iteration the agent is asked to select a word from a list of actions and write it.

Select randomly a word from the following list and write it.
Do not write additional text.

## START INPUT
{actions}
## END INPUT

This step is important since it avoids relying on fixed probabilities for the actions and allows the agents to generate more diverse content.

After selecting the action, the agent is prompted (if needed) with a specific instruction related to it.

Posts Generation

The agents generate posts by responding to the following prompt:

Write a short post introducing a topic of interest to you.

- Be consistent with your Big Five personality traits.
- Avoid excessive politeness.
- Do not exceed the limit. Make it short.
- Write in {language}.

Comments Generation

The agents generate comments by responding to the following prompt:

Read the following conversation and add your contribution to it.

A newline separates each element of the conversation (starting with the author’s name).
- You can tag the author of the tweet using @.
- Be consistent with your Big Five personality traits.
- Avoid excessive politeness.
- Your comment MUST contribute to the conversation.
- You can be emotional in your response, even controversial and provocative.
- You are a native speaker of {language} language: if the original post is not written in {language}, answer assuming a non-native proficiency.

## START CONVERSATION
{conversation}
## END CONVERSATION

Reactions Generation

The agents generate reactions by responding to the following prompt:

Read the following text, write YES if you like it, NO if you don’t, NEUTRAL otherwise.

## START TEXT
{post_text}
## END TEXT

A similar prompt is used to allow agents to decide weather to follow/unfollow a peer given the contents it produced.

News Sharing

The agents share news by responding to the following prompt:

Read the title and summary of the following article and share your thoughts about it.

- Be consistent with your Big Five personality traits.
- Avoid excessive politeness.
- Do not exceed the limit. Make it short.

## START INPUT
Title: {article.title}
Summary: {article.summary}
## END INPUT