Yager, Kevin G. (2024). Towards a science exocortex. Digital Discovery, 3(10). https://doi.org/10.1039/D4DD00175K
Yager outlines a vision for augmenting scientific reasoning through AI-powered tools capable of structuring, critiquing, and accelerating research workflows. This repository serves as a proof of concept for that vision, operationalizing a system where a swarm of AI agents engage in deliberative reasoning. Specifically, every agent invoked in support of a hypothesis is paired with a counter-agent tasked with critically evaluating its claims, ensuring that conclusions arise through structured opposition rather than passive agreement.
This repo, is a step towards that idea.
This prototype demonstrates the concept through a pipeline of reasoning agents. A support agent searches the literature for evidence in favor of a given hypothesis. An oppose agent then inverts the query to uncover contradictory or skeptical perspectives. Finally, a synthesis agent compares both outputs, identifies conflicts and research gaps, and proposes future directions or experiments to resolve ambiguity.
This deliberative pattern can be extended across the entire agentic swarm—for example:
- A data agent may construct both primary and inverse queries to identify bias in databases.
- An autonomous lab executor could test alternate procedures to ensure robustness in physical experimentation.
By embedding structured opposition into multi-agent workflows, this system moves towards the realization of a critical, reflective, and self-aware scientific exocortex.
.
├── Agents/
│ └── research_agent.py # Defines 3 agents: support, oppose, synthesize
├── Tasks/
│ └── research_task.py # Defines 3 tasks matching the agent flow
├── Tools/
│ ├── publication_tool.py # Fetches papers via arXiv API
│ └── embed_papers.py # Embeds paper abstracts (uses fastembed)
├── main.py # Entrypoint
├── crew.py # Defines agents + tasks into a crew
├── requirements.txt # Python dependencies
You can find sample outputs under Outputs/ folder for each task
You ask a question e.g. "Do foundation models show emergent reasoning?" The crew:
- support agent fetches supporting papers
- oppose agent fetches contradictory papers
- synthesize agent analyzes the conflict, assumptions, and blind spots
The system outputs:
- cited summary from each agent
- alignments, conflicts and gap for research
Getting Started
- Clone the repo
git clone https://github.com/SnehaDharne/exocortex.git
cd exocortex
- Set up a virtual environment
python3 -m venv venv
source venv/bin/activate- Install dependencies
pip install -r requirements.txt- Set your API key (for Gemini or OpenAI) Create a .env file:
GEMINI_API_KEY=your_key_here- Run the main script
python main.py