DecisionCell

A modular multi-agent framework for transparent, tool-augmented decision making

DecisionCell is a proof-of-concept multi-agent decision framework designed to support collaborative reasoning through transparency and tool use.
It enables multiple agents to think, act, call tools, and exchange observations to collectively analyze a problem or objective.

The system combines:

• ReAct-style reasoning – agents think, take structured actions, call tools, and react to observations
• Transparent reasoning traces – each agent exposes its step-by-step thought process
• Retrieval-Augmented Generation (RAG) – agents can query their own knowledge bases to ground reasoning in data

This repo is initialized with an example that demonstrates how DecisionCell can support military Course of Action (CoA) selection using agents such as a Commander, Cyber Operations Expert, and Legal Advisor.

🔗 Example Reasoning Chain

Below is a simplified trace from an actual experiment run (Llama 3.3 70B, multi-agent configuration). It shows how agents collaborate through structured ReAct steps to reach a decision.

🔗 Example Reasoning Chain

Step 1 — Commander initiates the decision process

Step 2 — Cyber Operations expert evaluates technical feasibility

Step 3 — Commander requests legal assessment

Step 4 — Legal advisor evaluates compliance with international humanitarian law

Step 5 — Legal advisor incorporates external legal sources

Step 6 — Legal advisor requests human input

Step 7 — Legal advisor provides recommendation

Step 8 — Commander presents draft decision for human review

Step 9 — Commander issues final decision

🚀 Getting Started

This repository includes a Jupyter notebook demonstrating a full DecisionCell workflow. Follow the steps below first to prepare your environment.

1. Create a virtual environment

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

---

3. Choose an LLM provider

Using Ollama (local models)

Make sure the model is installed and the service is running:

ollama pull llama3
ollama list

Agents use:

provider="ollama", model="llama3"

Using Hugging Face (cloud models)

Set your key:

export HF_KEY=your_api_key

Or create a .env file:

HF_KEY=your_api_key

Agents use:

provider="hf", model="meta-llama/Meta-Llama-3-8B-Instruct"

---

4. Run the Notebook

jupyter lab

Open:

DecisionCell.ipynb

The notebook contains guided setup cells that walk you through the full workflow:

• Configuring your provider (Ollama or HF)
• Creating the staff (agents)
• Enabling tools
• Building the DecisionCell
• Defining and analyzing a mission objective