MBTIverse: Simulating News Reactions with Agents

MBTIverse: Simulating News Reactions with Agents

Comment

Inspiration

Reanoning is a great functionality of LLMs that can be combined with generative technology to stimulate how humans think and react. Its always a challenge to see how people react to news for many people such as governments, politics, news agencies and social media creators and psychologists so they can plan how to handle things accordingly. With the help pf MBTI personalities which is one of the common and valid psycological way to understand human charactristics, I designed this platform so each agent can react to the input news as a person with that personality type does. I also implemented a meta reviewer agent that analyzes all reactions and reasnoning behind them and then proposes all useful information that can help user understand its society's reaction better.

In this project we considered all mbti types equally in final review and analysis, However, in real studies each country has a different population of each mbti type. Applying particular weight for each agent can stimulate a more realistic view of society reacting to the news. Moreover, in a wider view each society or country can be classified into a particular mbti type so globallly, this can also give a good overview of how each society or country react to news too.

How I Built It

  • Backend: FastAPI with a Swagger interface (no custom frontend).
  • GPT-OSS-120b Calls: Each MBTI agent loads its persona prompt from the prompts/ folder, using an OpenAI-compatible API.
  • Meta Reviewer: Runs with a higher token budget and lower temperature, producing structured JSON with sections like Consensus, Disagreements, Risks, Opportunities, and Recommended Actions.
  • Societal Impact Mapping: Each response is analyzed across impact domains with likelihood and severity scores.

The pipeline looks like this:

graph TD
    A[News Input] --> B[16 MBTI Agents]
    B --> C[Meta Reviewer]
    C --> D[Structured JSON Output]
    D --> E[Swagger UI Display]

## Challenges
- Harware requirement for GPT-OSS - Its very hard to run this model locally and even on a GPU eventhough there are several solutions available! I finally used huggingface APIs. (Its possible to run it on a GPU but its verrry slow!)
- Token budgeting — handling 16 agents’ outputs plus synthesis in one request.
- Psychological review is needed ( 16 agents' output)
- Meta reviewer is need to be reviewed by a social analyst too!
- Balancing structure and reasoning — agents needed to show why their MBTI type would react in a certain way, while keeping outputs parseable.

Built With

Share this project:

Updates