The original post explaning the solution Blog Post
div align="center">
The original GPTeam uses GPT-4 to create multiple agents who collaborate to achieve predefined goals. Their main objective ofwas to explore the potential of GPT models in enhancing multi-agent productivity and effective communication.
We appropriated the project to do research on deliberation over co-designing prompts and follow the simulations in a group setting, comparing real and simulated discussions, alsoto enact the issues and research possibility of catharsis in decision making.
Related papers:
Governance in Silico: Experimental Sandbox for Policymaking over AI Agents
https://dl.designresearchsociety.org/drs-conference-papers/drs2024/researchpapers/11/
Sandboxes as “trading zones" for engaging with AI regulation, ethics, and the EU AI Act: How to Reclaim Agency over the Future? https://osf.io/preprints/socarxiv/59qna
Video demo of the AI fable (appropriated the original use case): https://www.youtube.com/watch?v=cIxhI1d6NsM Soon more...
Read more about the architecture here: https://blog.langchain.dev/gpteam-a-multi-agent-simulation/
- install VScode - while not a must, it helps.
- install python 3.10+ - scroll down and pick you OS
- install git
If on windows, you will need to enable linux subsystem (WSL, here's an how-to)
and then inside VSCode click the blue box[1] and choose WSL[2]
inside vscode, run the following in the terminal
Clone the project repository to your local machine
git clone https://github.com/anonette/GPTeam
Move to the repository directory
cd gpteam
Run python setup.py to check your environment setup and configure it as needed
Update the environment variables in .env with your API Keys. You will need an OpenAI API key, which you can obtain here. Supplying API keys for optional services will enable the use of other tools.
Launch the world by running
poetry run world
To run the world cheaply, you can use poetry run world --turbo. This will use gpt3.5-turbo for all LLM calls which is a lot cheaper, but expect worse results!
Now you can observe the world in action and watch as the agents interact with each other, working together to accomplish their assigned directives.
GPTeam employs separate agents, each equipped with a memory, that interact with one another using communication as a tool. The implementation of agent memory and reflection is inspired by this research paper. Agents move around the world and perform tasks in different locations, depending on what they are doing and where other agents are located. They can speak to eachother and collaborate on tasks, working in parallel towards common goals.
The world is a busy place! To get a view of what different agents are doing whilst the world is running, you can visit the agents/ folder where there is a txt file for each agent containing a summary of their current state.
To change the world, all you need to do is:
- Make changes to the
config.jsonby updating the available agents or locations - Reset your database:
poetry run db-reset - Run the world again:
poetry run world
Read through the dedicated Discord setup docs
Make sure you have an ANTHROPIC_API_KEY in your env, then you can use poetry run world --claude which will run the world using claude-v1 for some calls and claude-v1-instant for others.
Make sure you have the Window extension installed, then you can use poetry run world --window. Some models may be slow to respond, since the prompts are very long.
We enthusiastically welcome contributions to GPTeam! To contribute, please follow these steps:
- Fork the project repository to your own account
- Create a new branch for your changes
- Implement your changes to the project code
- Submit a pull request to the main project repository
We will review your pull request and provide feedback as necessary.
Licensed under the MIT license.