# Colosseum

A research framework for auditing collusion between cooperative, tool-calling LLM agents, built on the Terrarium multi-agent environment, that identifies a taxonomy of collusion types across different models, network topologies and environments.

Colosseum is an experiment framework for objectively auditing collusion in cooperative multi-agent systems made of tool-calling LLM agents. It builds on the Terrarium (terrarium-agents) environment framework and provides a taxonomy of collusion types the authors study across varying models, agent-network topologies and task environments, with experiment code living under an experiments directory.

Some environments require the separate CoLLAB library, installed either as a git submodule or cloned separately and referenced through an environment variable. Setup uses a uv-managed Python 3.11 virtual environment, and API keys for whichever model providers are used go in a local .env file; the authors warn that API costs scale with the number of agents in a run.

It is a research artifact accompanying a paper rather than a general-purpose product, intended for reproducing or extending experiments on multi-agent collusion.

Website: https://github.com/umass-ai-safety/colosseum