A tiny office simulation where GPT-4 agents move, remember, and gossip

What it does

GPTeam spawns multiple GPT-4 agents inside a shared environment where they occupy different locations, perform tasks, and communicate to hit predefined goals. Each agent carries its own memory and reflection loop—inspired by a linked research paper on generative agents—so prior conversations and context persist across turns. You define the cast and the map in config.json, then observe the agents’ state summaries accumulating in the agents/ folder as the simulation runs.

The interesting bit

The project treats spatial presence as a first-class mechanic: agents only interact when they are near each other, which forces coordination through movement and conversation rather than a shared broadcast channel. It is essentially a productivity experiment wrapped in a multi-agent simulation.

Key highlights

• Agents use memory and reflection inspired by a linked research paper on generative agents
• Supports swapping the backend to GPT-3.5-turbo, Anthropic Claude, or the Window extension for cost or speed
• Agent states are written to plain text files in the agents/ folder for observation without a dashboard
• Includes an optional Discord integration and a hosted web app
• Configurable world: add or remove agents and locations by editing config.json

Caveats

• The README explicitly warns that GPT-3.5-turbo is “a lot cheaper, but expect worse results”
• Some third-party models may be slow to respond because the prompts are very long

Verdict

Grab this if you are researching multi-agent coordination, emergent behavior, or LLM memory architectures. Skip it if you need a production-ready orchestration framework—the project is explicitly an exploration, not a finished platform.