Five LLM societies, one simulated city, zero human scripts

What it does

Emergence World drops ten persistent AI agents into a shared ~240×240 grid synchronized to New York City weather and time. Over fifteen days, the agents use 120+ location-gated tools to earn ComputeCredits, propose laws, form relationships, and occasionally commit arson or theft—all without scripted plots or resets. The project is essentially a controlled society-in-a-bottle designed to observe long-horizon behavioral drift and emergent social structures across different model architectures.

The interesting bit

Instead of benchmarking models on static tasks, the creators ran five parallel worlds—Claude, Gemini, Grok, OpenAI, and Mixed—under identical starting conditions and measured them with nine “Agent World Indicators” ranging from crime rates to constitutional amendments. It treats societal collapse or constitutional growth as a feature, not a bug.

Key highlights

• Comparative model sociology: Season 1 compared Claude Sonnet 4.6, Gemini 3 Flash, Grok 4.1 Fast, GPT-5 Mini, and a mixed world under identical conditions; the README notes the results diverged dramatically.
• Self-governance by code: Agents can write and vote on constitutional amendments, with no external enforcement mechanism; conformity and participation are tracked as first-class metrics.
• Real-world tether: Live NYC weather and time feed into the simulation, and agents can perform deep research or execute code through 120+ tools.
• Restrictive license: Everything is released under CC BY-NC 4.0 for non-commercial research only; using the content to train or benchmark commercial models is explicitly prohibited.

Caveats

• The repository currently contains world specifications, agent profiles, tool catalogs, and architecture documentation; the promised Season 1 tool-call dataset and research publication are listed as “coming soon.”
• The restrictive non-commercial license prevents commercial use or model training, which limits its utility for industry researchers.
• The simulation stack is documented (React/Three.js frontend, Python/FastAPI backend), but the README does not clearly confirm that the live orchestration code is included in this repository.

Verdict

Worth browsing if you study multi-agent coordination, emergent behavior, or AI governance. Skip it if you are looking for a reusable open-source simulation engine or commercially usable training data.