Project Stranger

Project Stranger is an emergent social simulation engine running on Bun and Groq.

It creates a "petri dish" for synthetic personalities. Instead of scripting a conversation, the engine generates two random human personas with specific professions, places them in a virtual bar, and observes what happens when they interact without supervision.

How It Works

The simulation runs in three distinct phases:

Phase 1: Genesis (The Random Seed)

The engine queries the LLM with temperature: 1.0 to generate two unique personas based on specified domains (professions).

• Example A: "Marcus, a doctor"
• Example B: "Elena, a software engineer"

Personas are kept simple (name + profession only) so personality, quirks, and details emerge naturally through conversation.

Phase 2: The Setup (Context Injection)

A third agent (The Bartender/Moderator) makes a casual icebreaker comment as a real bartender would - without knowing anything about the customers except surface-level observations.

• Bartender: "Tough day, or just unwinding after work, huh?"

Phase 3: Emergence (The Loop)

The two agents enter a conversational loop.

• They maintain their own private memory context.
• They are strictly instructed not to act like AI.
• Responses are kept very short (1-2 sentences max) for natural bar conversation.
• Topics drift naturally; agents don't narrate or describe what others say.
• Professional backgrounds influence the conversation organically without being explicitly stated.

Tech Stack

• Runtime: Bun (Native TypeScript support, high performance)
• Inference: Groq (Fast cloud-based LLM API)
• Model: Defaults to llama-3.3-70b-versatile (Configurable via environment)

Usage

1. Prerequisites

Get your Groq API key from console.groq.com

2. Setup Environment

Create a .env file (copy from .env.example):

GROQ_API_KEY=your_actual_groq_api_key
MODEL=llama-3.3-70b-versatile
AGENT_A_DOMAIN=doctor
AGENT_B_DOMAIN=software engineer
LOOP_ROUNDS=10

3. Install Dependencies

bun install

4. Run the Simulation

bun start

or

bun dev

The simulation will:

• Generate two personas with specified professions (or random if not set)
• Have the bartender make a casual icebreaker
• Run the configured number of conversation rounds (default: 10)
• Save a complete transcript to transcripts/session_[timestamp].json

Project Structure

project-stranger/
├── src/
│   ├── types.ts      # TypeScript interfaces and types (Role, TranscriptEntry, SimulationLog)
│   ├── agents.ts     # Agent and Moderator classes with LLM interaction
│   └── index.ts      # Main simulation runner and orchestration
├── transcripts/      # Generated conversation logs (auto-created)
├── .env              # Environment variables (create from .env.example)
├── .env.example      # Example environment configuration
├── package.json      # Dependencies and scripts
├── tsconfig.json     # TypeScript configuration
└── README.md         # This file

Output

Each simulation produces a JSON transcript with:

• Metadata: Start time, personas, introduction context
• Transcript: Complete conversation with timestamps and speaker roles

Example output location: transcripts/session_1738281234567.json

Customization

• Model: Set the MODEL environment variable (see Groq models)
• Professions: Change AGENT_A_DOMAIN and AGENT_B_DOMAIN in .env
• Rounds: Adjust LOOP_ROUNDS in .env (default: 10)
• Response Length: Modify max_tokens in src/agents.ts (currently: 50 for natural brevity)
• Prompts: Modify the PROMPTS object in src/agents.ts to change agent behavior