GPTeam: Collaborative AI Agents
An open-source multi-agent simulation where several GPT-4 agents, each with its own memory, collaborate in a simulated world to work toward shared goals.
GPTeam uses GPT-4 to run multiple agents that collaborate to reach predefined goals, exploring how language models can improve multi-agent productivity and communication. Each agent has its own memory and reflection process, inspired by prior generative-agent research, and agents move around a simulated world, perform tasks at different locations, and talk to each other to coordinate work.
The world and its agents are configured through a JSON file that defines available agents and locations, and the project can be run more cheaply with GPT-3.5-turbo using a turbo flag, at the cost of weaker results. It also supports Anthropic's Claude as the underlying model, and includes a Discord integration and a per-agent state summary written to a text file so a user can watch what each agent is doing. It is written in Python and licensed under MIT.