ExoNet: Idea Illustration & Feasiblity

Overview

This simple multi-agent framework in chat_multi.py offers collaborative AI agents with their own specialized backgrounds/expertise and flexible interaction order!

Citation: Tsai, Esther HR, and Kevin G. Yager. "Exocortex Network for AI-Augmented Human-Led Scientific Expedition." Photon Science (2025). https://doi.org/10.1021/photonsci.5c00009

Key Features

🎯 Individual Agent Backgrounds

• Specialty Definitions: Clear role and expertise areas for each agent
• Model Configuration: Individual model settings per agent
• PDFs: Each agent can have access to specific PDF documents
• Images: Visual context through image files
• Websites: Web content as background knowledge
• Text Files: Additional text-based context

🔄 Flexible Interaction Patterns

• Custom Interaction Order: Define how agents collaborate, e.g., [1, 3, 2, 4]
• Multiple Rounds: Configurable deliberation cycles. Each round contains the entire list of interaction order. For example, two rounds would be [1, 3, 2, 4] and [1, 3, 2, 4]
• Collaborative Learning: Agents build upon each other's responses
• Human interaction: After running default_question for the specified number of rounds, user can prompt new questions.

Usage

Command Line Interface

# Basic usage with config file
python chat_multi.py --config config_example.json

# Enable verbose output
python chat_multi.py --config config_example.json --verbose

Example Configuration

{
  "title": "Enhanced Multi-Agent System Demo",
  "description": "Multi-agent system with individual backgrounds, specialties, and flexible interaction order",
  "max_tokens": 128000,
  "rounds": 5,
  "clean_slate": true,
  "interaction_order": [1, 2, 1],
  "default_question": "Which is more important in DOE national laboratories, choose one: (A) data and AI security, (B) develop and deploy AI agents for science",
  "agents": [
    {
      "name": "Scientist",
      "model": "gpt-5",
      "specialty": "You are a senior chemist with also expertise in computer science, You are a chemist with also expertise in computer science, allowing you to be visionary and bridge the gap between physical science and computational science/AI. Provide a clear, concise proposal to the user's question. When provided, incorporate the attached document/image.",
      "char_limit": 3000,
      "websites": [ 
        {
            "url": "https://www.bnl.gov/staff/kyager",
            "name": "Staff",
            "crawl_depth": 1,
            "max_crawl_pages": 1
        },   
        {
            "url": "http://yager-research.ca/",
            "name": "KY-research",
            "crawl_depth": 1,
            "max_crawl_pages": 10
        }
        ],
      "pdfs": ["data/KY/Yager_Exocortex.pdf"]
    },
    {
      "name": "Scientist",
      "model": "gpt-5",
      "specialty": "You are a computational scientist with also expertise in economics. Provide a clear, concise proposal to the user's question. When provided, incorporate the attached document/image.",
      "websites": [ 
        {
            "url": "https://www.bnl.gov/staff/etsai",
            "name": "Staff",
            "crawl_depth": 1,
            "max_crawl_pages": 3
        }
        ],
      "images": ["data/images/example.jpg"],
      "text_files": ["data/texts/example1.txt", "data/texts/example2.txt"]
    }
  ],
}

Suggestions

• Azure OpenAI: The implementation employs direct calls via Azure OpenAI: API key, base, deployment name should be specified accordingly. See line 370 in chat_multi.py and specify accordingly in ~/.bashrc e.g. export AZURE_API_KEY=XXX
• max_tokens: Higher max_tokens (6000-8000) for detailed analysis, while lower max_tokens for concise responses. However note that if too constrained, model may have trouble outputting reasonable response.
• char_limit For background information, web crawling, char_limit (e.g. 2000) should be chosen depending on the number of documents and the context window.
• clean_slate: Starting fresh in every round allows for convenient repetition of the question across multiple rounds, acquiring statistical information to account for the model stochasticity.