XiaoLuoLYG/GOD
Govern, Observe, Direct - a real-time control room for agent societies
What it solves
GOD provides a control-centric environment for inspecting and manipulating language-agent societies. Unlike traditional agent simulations that are passive observations, GOD allows operators to actively intervene, question agents mid-run, and precisely control the timeline of the simulation to understand how agents respond to specific changes or instructions.
How it works
GOD operates as a local-first system consisting of a React/Vite frontend (the Control Room), a FastAPI backend, and an out-of-process LLM agent runtime. It separates the simulation into three core concepts: Experiments (the setup, including maps and agent profiles), Replays (the recorded results), and Runtime State (local machine data). Users can configure the system via a browser-based no-code setup wizard that handles model configuration and scenario generation.
Who it’s for
It is designed for researchers and developers who want to inspect the behavior of LLM-based agent societies, test hypotheses about agent interactions, and create reproducible agent-based social simulations.
Highlights
- Replay Control: Ability to pause, scrub, fast-forward, and auto-play simulation steps.
- Targeted Intervention: Inject instructions into the next step of the simulation to steer agent behavior.
- Direct Communication: Ask natural-language questions to individual agents, groups, or the entire town mid-simulation.
- No-Code Setup: A browser wizard for configuring models, generating agent profiles, and designing scenarios without editing code.
- Studio Tools: Dedicated Agent and Map Studios for creating and calibrating custom residents and world maps.
- Portable Packs: Import and export experiments, maps, and agent packs via ZIP files.