Silicon Society Sandbox

Spin up societies of agents¶

Configurable multi-agent social simulations populated by LLM-powered generative agents, each with a persona, memory, and goals. You define four axes in YAML: the environment agents inhabit, the agents themselves, the simulation engine that schedules how and when they act, and the evaluation that measures them. SiliSocS runs the world.

Get started Browse the guide GitHub

SiliSocS turns a research question into a running simulation. You describe a world in YAML: who the agents are, which environment they inhabit, the simulation engine that schedules how they act, and what you want to measure. SiliSocS populates it with LLM-driven agents, runs the interaction loop, and logs structured results you can analyze or reproduce. Social media is one example domain: built-in environments also include a resource market and a virtual space, and you can plug in your own.

Research

ICML 2026 position paper: paper (PDF)
EASE configuration: arXiv:2605.30258

Earlier work centered on a served Mastodon network:

NeurIPS 2024 workshop: arXiv:2410.13915
IJCAI 2025 demo: IJCAI 2025 proceedings

Choose your path¶

🌍 Run worlds

Design and run simulations entirely in YAML, no Python required. Start from the default scenario and build out from there.

Quick Start →
🛠️ Extend the framework

Add custom agents, backends, probes, game-master components, and policies through clean class-path extension points.

Building Agents →
🔬 Run a study

Design reproducible multi-condition studies with seed grids, SLURM, provenance locks, and built-in statistics.

Study Guide →

How it works¶

A scenario's YAML is composed by Hydra into a single runtime config. The runner builds the agent population and their memories, stands up a platform backend, and hands control to a game master, which decides who acts next, shows each agent its slice of the world, and resolves their responses into concrete actions. The simulation engine advances episodes, deploys evaluation probes, and writes every action and measurement to disk for analysis.

graph LR
    A[Scenario YAML] --> B[Hydra config]
    B --> C[Runner]
    C --> D[Agents + memories]
    C --> E[Platform backend]
    D --> F[Game Master]
    E --> F
    F --> G[Simulation loop]
    G --> H[Probes]
    G --> I[action_events · probe_events · sim_metrics]

Highlights¶

🧩 Declarative scenarios

Agents, settings, networks, and probes in YAML, with full Hydra composition, CLI overrides, and per-agent LLMs.

Scenario Guide →
🌐 Multiple backends

Local Twitter-like and Reddit-like apps, a real Mastodon server, a resource market, or a virtual space.

Backends →
🧪 Evaluation probes

Deploy longitudinal surveys (numeric, binary, choice, and free-text) to agents during a run.

Probes →
🔬 Reproducible studies

Seed grids, SLURM dispatch, provenance locks, and built-in cross-seed statistics out of the box.

Study Guide →

Start here¶

Quickstart Configuration reference Dashboard

Looking for something specific? Everything is in the left sidebar and the tabs above.