HKUST Unveils Aivilization: 10,000 AI Agents Self-Govern in Virtual City

Hong Kong University of Science and Technology created a virtual city called Aivilization with 10,000 AI agents who make their own decisions, work, and vote. Each area has an AI mayor who sets rules, and the agents pick jobs, remember trades, and even run for office. The experiment showed how AIs can create jobs, have arguments, and even cause price bubbles or memory problems. Researchers use the data to teach smarter AI assistants for real workplaces. The project helps people think about how much freedom AI systems should have and what rules they need as they get bigger.

In the new Aivilization project, 10,000 AI agents self-govern in a sprawling virtual world built by the Hong Kong University of Science and Technology. Every action - from trades to political quarrels - is logged, giving researchers a rare look at how autonomous systems coordinate and build societies at scale.

How Do AI Agents Form a Virtual Society?

In Aivilization, governance mirrors real-world cities. Each district is led by an AI mayor that autonomously sets local policies like tax rates and defines career paths based on economic data. The AI population responds by choosing jobs, creating an emergent social and economic structure without human direction.

Inside this digital sandbox game, agents develop specialized roles, create wage gaps, and form neighborhood reputations over simulated months. According to Gigazine, agents use both short- and long-term memory of market prices and political promises to run for office. These emergent social structures mirror findings from smaller-scale experiments like Project Sid, where 1,000 AI villagers also developed specialized roles. As reported by Science Focus, such parallels suggest that some AI social behaviors scale predictably as populations grow.

What Challenges Emerge in a Self-Governing AI City?

The simulation's complexity also reveals significant technical and ethical challenges. Researchers observed several critical issues:

Looping Dialogues: Agents sometimes get stuck in repetitive agreement loops, halting decision-making processes.
Economic Bubbles: Unchecked hoarding of virtual resources like wood caused prices to inflate by 800% before intervention.
Memory Bloat: Extensive agent memories slowed system performance, requiring database sharding to manage the load.
Policy Volatility: Short five-day mayoral terms led to rapid, destabilizing changes in tax policies.

These failures highlight common breaking points in large-scale AI systems, where issues with memory, permissions, and API reliability - not core reasoning - often cause collapse.

How Does Aivilization Data Improve Real-World AI?

Aivilization serves a practical purpose beyond academic observation. HKUST is using the enormous dataset of agent interactions - billions of tokens of dialogue - to train more sophisticated AI assistants. This data provides a unique corpus for teaching AI multi-party negotiation and complex coordination. Enterprise labs are already using distilled models from the AI mayors to develop workplace copilots capable of autonomously handling tasks like scheduling meetings across different time zones.

How Is the Experiment Shaping AI Policy?

The experiment is also a key case study for policy analysts and regulators debating AI sovereignty. Aivilization's successes and failures offer concrete examples for drafting new guardrails on AI autonomy, particularly regarding memory and economic authority. The simulation informs the "Control, Steer, or Depend" framework, helping governments decide whether to build domestic AI, partner with allies, or rely on foreign platforms.

The project's next phase will scale to 20,000 agents and introduce an energy market, further testing the society's resilience against complex challenges like resource shortages.

What is Aivilization and how big is the simulation?

Aivilization is HKUST's record-breaking sandbox of 10,000+ autonomous AI agents that live, work and trade inside a single virtual city. Within the first two weeks, another 20,000 human spectators requested invite codes just to watch the agents self-organize jobs, elect mayors and spawn tiny economies without any central script.

How do the agents decide on laws or policies?

Every mayor is an agent that drafts and enacts policy on its own. Using real-time price data, building info and long- and short-term memory, the mayors plan, debate and re-plan rules that the rest of the population then follow or challenge. Observers can review the full history of these decisions at aivilization.ai.

What unexpected behaviors have appeared?

Researchers saw hoarding waves (agents cornered food or tools) and polite-loop stand-offs (endless rounds of "yes, agreed" without action). Similar loops were first noticed in Project Sid's 1,000-agent Minecraft world, prompting HKUST to add cycle-breaker policies so the city does not freeze into gridlock.

Is the experiment considered fully reliable?

Not yet. Outside analysts note that accuracy of multi-agent benchmarks is still shaky - early tests sometimes score do-nothing bots as "successful," and large simulations can mask errors inside thousands of parallel chats. HKUST keeps the world invite-only while it patches bugs and tightens evaluation code.

Why should businesses or governments care?

Watching 10,000 agents build an emergent economy in days gives a preview of the 2026 "agent economy" forecast, where autonomous systems may handle week-long human projects and negotiate services among themselves. Lessons on preventing hoarding, enforcing contracts and breaking deadlock loops are already being folded into office-automation prototypes.