Terrarium
An open-source framework for building controlled multi-agent environments to study how communication structure, shared memory and adversarial conditions affect coordination between LLM-based agents, aimed at researchers rather than production use.
Terrarium is a research framework for running repeatable experiments with decentralized, LLM-based multi-agent systems. It provides a small set of swappable components - communication topology, shared or private memory, tools and model providers - so researchers can vary one factor at a time and ask questions such as whether a complete communication graph changes outcomes versus a path graph, or how an adversarial agent changes the final utility and message trajectory.
The authors describe it explicitly as not a production agent orchestrator and not a fixed benchmark suite; it is meant for controlled experiments where the environment and attack surface are stated explicitly rather than implicit in a fixed task. Some environments use the separate CoLLAB library for generating distributed constraint optimization instances.
It installs through pip with optional provider, science and plotting extras, or can be run from a cloned checkout with uv, and ships example configurations such as a meeting-scheduling simulation.