deb8flow is a full-stack, multi-agent AI debate framework built with LangGraph. It simulates a structured and fact-checked formal debate between two autonomous AI agents — one PRO and one CON — orchestrated by a moderator and finalized with a judge's verdict.
This project showcases complex agentic flows, runtime control logic, fact verification, and dynamic decision-making in a fully controlled LangGraph workflow. If you're exploring how to build serious multi-agent systems with memory, routing, retries, and arbitration, deb8flow is your guide.
-
Dynamic Multi-Agent Debate Flow
PRO vs CON debaters exchange structured arguments, rebuttals, counters, and final remarks. -
LLM-Powered Agents
Built using OpenAI’s GPT-4.1 (with web search) and LangChain runtimes. -
Fact Checker Agent
Validates claims using OpenAI's integrated web search; enforces accuracy with retry logic and disqualification if falsehoods persist. -
Moderator & Router Nodes
UsesCommandobjects to orchestrate debate flow and enforce stage transitions. -
Judge Agent
Concludes the debate with a verdict based on clarity, structure, and persuasive merit. -
Fully Modular LangGraph Architecture
Clean separation of nodes: topic generator, debaters, fact checker, routers, and judge.
- Topic Generator: Produces a nuanced, debatable topic for the session.
- Pro Agent: Delivers the opening statement in support of the topic.
- Moderator: Passes the floor to the Con Agent.
- Con Agent: Responds with a rebuttal to the Pro’s opening statement.
- Moderator: Returns the floor to the Pro Agent.
- Pro Agent: Offers a counterargument to the Con Agent's rebuttal.
- Moderator: Passes the floor to the Con Agent again.
- Con Agent: Presents their final closing statement.
- Moderator: Wraps up and hands off to the Judge Agent.
- After every speaker, the Fact Checker evaluates each claim for accuracy.
- If a statement fails the fact-check, the speaker must restate or correct their argument.
- Any debater who fails three fact-checks automatically loses the debate.
- The Judge Agent reviews the full debate history and evaluates both sides based on:
- Argument quality
- Logical clarity
- Structural coherence
- The most persuasive, well-reasoned side wins.
- Clone the repository:
git clone https://github.com/iason-solomos/Deb8flow.git
cd deb8flow
- Install Dependencies
pip install -r requirements.txt
- Set up your .env file
Should include those keys:
OPENAI_API_KEY
and optionally (but recommended) for tracing:
LANGCHAIN_API_KEY
LANGCHAIN_TRACING_V2
LANGCHAIN_PROJECT
- Run the debate
python main.py
deb8flow/
├──configurations/ # Debate constants and LLM configs
├──nodes/ # LangGraph nodes (debater, fact checker, etc.)
├──prompts/ # System and human prompts per agent
├──tests/ # Tests
├──workflow/ # Langgraph workflow
├──debate_state.py # Typed state definitions (DebateState, DebateMessage)
├──main.py # Entry point to run the debate
├──README.md
├──requirements.txt
├──utils.py # Some common utilities like getting the debate history
└──workflow_graph.pngOpenAI GPT-4.1
Python 3.10+
Pydantic
dotenv
Argument Generation or Position Testing
Fact Checking + Self-Correction Pipelines
Multi-Agent Workflow Demos
LangGraph Pattern Reference
AI Safety / Alignment Experiments
Add scoring by a panel of judges and decide winner based on majority voting
Add multiple debate rounds or crossfire
Export debate timeline to Markdown or HTML
Want to add a new node, logic layer, or feature? PRs welcome!
Create your branch: git checkout -b feature/your-feature
Commit your changes
Push and open a pull request
MIT License. © 2025 [Iason Solomos]
If you find this useful, star the repo ⭐ and share it with your LangChain, LLM, or AI dev friends!