AIScientists-Dev/WorldSeed

What it solves

WorldSeed provides a way to create complex, multi-agent simulations without building rigid workflows. Instead of scripting every interaction, it allows users to "seed" a world by defining rules, roles, and consequences, letting emergent behaviors and useful artifacts arise naturally from agent interactions.

How it works

The system operates on a "tick loop" where the world advances one step at a time. In each tick, agents perceive a filtered slice of the world based on asymmetric information rules, propose actions, and the engine resolves them. Predictable outcomes are handled by a deterministic DSL (Domain Specific Language) engine, while uncertain outcomes are judged by an LLM-based "Dungeon Master" (DM) that returns structured effects. The entire world—including entities, physics, and perception—is declared in a single YAML configuration file.

Who it’s for

It is designed for developers and researchers building multi-agent worlds, whether for production simulations, AI-driven research communities, games, or fictional narratives.

Highlights

• Scene-Agnostic Engine: The same core engine can run vastly different scenarios, from an AI research lab to an espionage drama, simply by changing the YAML config.
• Asymmetric Information: Built-in perception rules ensure agents have different perspectives of the same event, mimicking real-world information gaps.
• Hybrid Resolution: Combines a fast, deterministic rule engine with an LLM-based DM for nuanced judgment.
• Interactive Observation: Users can watch the simulation from a high-level dashboard, intervene by whispering to agents, or step into a character to play alongside the AI.
• AI-Assisted World Creation: Includes a /create-world skill that generates the necessary YAML and UI configurations from a text prompt.