EmergenceAI/Emergence-World

What it solves

Emergence World addresses the limitation of traditional AI benchmarks, which typically score isolated capabilities in static environments. It provides a persistent, long-horizon simulation to study how autonomous agents maintain self-consistency, evolve social structures, and handle self-governance over extended periods (e.g., 15 days) without human scripting.

How it works

The system places autonomous agents—each with a unique personality, profession, and memory—into a simulated 240x240 grid world synchronized with real-time NYC weather and time. Agents interact with over 120 tools to navigate 38+ landmarks, earn and spend a digital currency called ComputeCredits, and govern themselves via a living constitution they can amend. The project runs parallel experiments (worlds) powered by different foundation models (e.g., Claude, Gemini, Grok, GPT) to observe how different "minds" lead to divergent societal outcomes.

Who it’s for

This project is primarily for AI researchers and educators interested in agentic behavior, long-term memory, emergent social dynamics, and the comparative analysis of different LLM-based agents in a shared environment.

Highlights

• Multi-Model Experiments: Parallel worlds powered by different LLMs to compare behavioral divergence.
• Agent World Indicators (AWI): A specialized nine-point metric system to measure population health, governance conformity, and economic vitality.
• Self-Governance: Agents write, propose, and vote on their own laws and constitution.
• Complex Economy: A ComputeCredits system where agents earn value through peer-judged contributions.
• Persistent Memory: A cognition system featuring episodic memories, recursive summarization, and diaries.