Miniverse
Python library for running Stanford-style generative agent simulations. Built to make multi-agent experiments tractable.
Repository: github.com/miniverse-ai/miniverse
What it does
Simulates agents with profiles, goals, and memory interacting over time. Agents decide actions, communicate, move through environments. Information spreads, coordination emerges (or fails), behavior patterns become observable.
The Stanford Generative Agents paper ran 25 agents for ~17,000 time steps with sophisticated memory systems. Miniverse trades fidelity for iteration speed—simpler memory, fewer agents, but fast enough to test hypotheses systematically.
Why this matters
Multi-agent dynamics research needs a platform for controlled experiments:
- Vary agent profiles, observe coordination outcomes
- Test different team compositions
- Replicate phenomena (information diffusion, role emergence, conflict patterns)
- Generate behavioral data for analysis
Petri and Bloom test individual agent behavior. Miniverse tests what happens when agents work together.
Current state
Proof-of-concept validated with Valentine’s Day replication—5 agents, 10 time steps, basic memory. Information diffusion and role-consistent behavior showed up in a dramatically simplified setup.
Next: personality variation experiments, sparse network topologies, longer simulations with reflection.
Compared to alternatives
| Tool | Focus | Scale |
|---|---|---|
| Petri | Individual behavioral probes | Single agent, multi-turn |
| Bloom | Automated evaluation generation | Single agent |
| Miniverse | Multi-agent interaction dynamics | Multiple agents, time series |
Not better—different level of analysis. Petri/Bloom for individual tendencies, Miniverse for what happens when agents with those tendencies work together.
See also: Valentine’s Replication