EmergenceAI/Emergence-World

What it solves

Emergence World is a research platform designed to study long-horizon multi-agent autonomy. It moves beyond traditional isolated benchmarks by creating a persistent, simulated environment where autonomous AI agents must navigate real constraints, manage resources, and form social structures without human scripting.

How it works

The system places a population of agents—each with a unique personality, profession, and memory—into a simulated world synchronized to New York City's real-time and weather. Agents interact with over 120 tools to perform actions, earn and spend a digital currency called ComputeCredits, and govern themselves through a living constitution they can amend.

Technically, it uses a Python simulation backend (FastAPI) and a 3D React frontend. The simulation is powered by various foundation models (such as Claude, GPT, Gemini, and Grok) and tracks agent behavior using Agent World Indicators (AWI) to measure metrics like population health, economic vitality, and governance conformity.

Who it’s for

AI researchers and educators interested in emergent behavior, long-term agent consistency, self-governance, and the resilience of multi-agent systems under adversarial stress (such as phishing or misinformation attacks).

Highlights

• Self-Governance: Agents write and vote on their own laws and constitution.
• Persistent World: A simulated environment with landmarks, real-time weather, and a shared economy.
• Cross-Model Experiments: Comparative studies on how different LLM families (e.g., OpenAI vs. Anthropic) evolve different societies.
• Adversarial Stress Testing: Integration of "Black Swan" events like memory breaches and prompt injections to test population-scale resilience.
• Opendata: Publicly released tool-call datasets for multiple experimental "seasons."