Society Simulator Agent
Inspiration
In today’s digital world, content goes viral—or fails—within minutes, yet creators have no reliable way to predict audience reaction before publishing.
We were inspired by a simple question:
What if you could simulate society before you speak to it?
Every post, reel, advertisement, or idea triggers diverse emotional responses across different segments of society. However, most systems only optimize for engagement after publication, not before.
This project was built to bridge that gap by designing an AI system that simulates human-like reactions using multi-agent reasoning.
What it does
Society Simulator Agent is a multi-agent AI system that predicts how different segments of society will react to any piece of content before it is published.
Given an input such as a social media post, advertisement, or idea, the system:
Simulates reactions from multiple personas:
- Gen Z student
- Working professional
- Meme/content creator
- Critical/skeptical user
- Neutral observer
- Gen Z student
Predicts:
- Engagement Score
- Virality Score
- Controversy Score
- Engagement Score
Uses MCP-integrated external data sources (Elastic / MongoDB) to retrieve historical viral content patterns.
Provides:
- Content improvement suggestions
- Optimized version of the input content
- Content improvement suggestions
In essence, it does not just analyze content — it simulates societal response before publication.
How we built it
We built the system as a multi-agent architecture powered by Gemini and Google Cloud Agent Builder, integrated with MCP-based data grounding.
Core Architecture
Planner Agent (Gemini)
- Breaks input into structured subtasks
- Identifies intent, audience, and tone
Persona Simulation Agents
- Generate independent reactions from societal segments
- Output sentiment, behavior, and reasoning
Trend Intelligence Agent (MCP Integration)
- Retrieves similar viral content using Elastic/MongoDB MCP servers
- Grounds predictions in real-world engagement patterns
Simulation Engine
- Aggregates all persona outputs
- Computes final scores
Optimization Agent
- Suggests improvements
- Generates optimized content
Tech Stack
- Google Cloud Agent Builder
- Gemini 3 (reasoning + planning)
- MCP (Elastic / MongoDB integration)
- FastAPI backend
- Lightweight frontend dashboard
Challenges we ran into
Designing a truly agentic system instead of a single LLM call was the biggest challenge.
We had to carefully structure multi-step reasoning flows, persona consistency, and output aggregation.
Another challenge was integrating MCP meaningfully—not just as a data connector, but as an active reasoning component influencing predictions.
Balancing creativity, consistency, and real-world grounding was also a key difficulty.
Accomplishments that we're proud of
- Built a fully multi-agent AI system instead of a chatbot
- Integrated MCP-based real-world trend intelligence
- Simulated societal behavior using structured personas
- Enabled automatic content optimization
- Designed a complete end-to-end agentic workflow
What we learned
We learned that agentic systems are fundamentally different from traditional AI applications.
Key insights:
- Orchestration is more important than raw model intelligence
- Clearly defined agent roles prevent reasoning collapse
- External grounding via MCP significantly improves reliability
- Simulation-based AI can outperform static prediction models
We also learned how powerful Gemini becomes when embedded inside structured execution pipelines.
What's next for Society Simulator Agent
We aim to evolve this into a full-scale pre-publication intelligence platform for creators, marketers, and brands.
Future improvements include:
- Real-time social media trend ingestion
- Advanced persona modeling with demographic tuning
- Emotional trajectory prediction over time
- Integration with platforms like Instagram, YouTube, and X
- Plugin system for marketing agencies
Final Vision
No content is published blindly ever again.
Built With
- api
- gemini
Log in or sign up for Devpost to join the conversation.