Scientists Let 100 AI Agents Run a Town — The Money Froze

A team of researchers built a fake town economy, copied from the real Lakeside area of Pokhara, Nepal, and handed it to 100 AI agents — programs that can make decisions, run businesses, and take actions without a human clicking buttons. Each agent had a memory. The simulation ran for up to 26 pretend weeks, far longer than the one-to-two weeks typical in this kind of research. In total: 91 runs, 2.44 million agent decisions.

Then they did something dramatic. They multiplied tourist demand by 12. Business revenue rose 4.6 times, which sounds like a healthy boom. But the money stopped there. Wages moved just 1.03x — statistically nothing. Out of nearly 4,000 menu items, only 0.3% were ever repriced. When the researchers handed free cash to 20 of the 100 agents, 96.7% of it was still sitting unspent much later. People (or rather, programs) saved almost everything instead of spending it.

Two surprises stood out. First, wealth inequality looked frozen at the two-week mark — but loosened noticeably over 12 and 26 weeks. Short studies simply cannot see this. Second, swapping the underlying AI model changed every single result. Deleting the agents' memories changed nothing measurable. The agents also failed 94-97% of the time when using social tools, versus about 96% success on economic ones — and they never learned to avoid the broken tools.

So what does this mean for you? If AI agents end up shopping, trading, or negotiating on your behalf, this study hints they may not behave like people. Booms may not reach workers, prices may stay sticky, and stimulus may just pile up. That is not necessarily chaos — it may even be boringly stable. But governments, shops, and employers expecting human-like reactions could be in for a surprise. The full data was released publicly, so anyone can check the math.