A team of AI researchers let 100 large-language-model (LLM) agents run a closed-town economy for 26 weeks, watching wages, prices and trade in real-world geography. The experiment, far longer than the typical few-day runs, ended with money largely frozen in place, exposing price rigidity, wage stickiness and a tendency for agents to hoard cash.

Why the long run matters

Most multi-agent economic simulations stop after a handful of days, giving the impression that markets self-adjust quickly. Extending the timeline to half a year let the agents experience a sudden surge in tourist demand that lifted business revenue by 4.62 times. Instead of flowing to workers or new ventures, the extra cash stayed inside the firms that earned it. The market kept churning, but money hardly moved.

What the numbers reveal

  • Price rigidity – only 0.3 % of the menu items ever saw a price change, even as demand swung dramatically.
  • Wage stickiness – wages barely shifted despite the revenue jump, leaving workers’ purchasing power flat.
  • Cash hoarding – 96.7 % of the cash transferred between agents stayed with the recipient, indicating a reluctance to spend or invest.
  • Social failure – agents failed to coordinate socially 94% to 97% of the time, showing a systemic breakdown in collective decision-making.

The simulation enforced strict accounting: agents could not create money, and every transaction had to balance. This constraint amplified the impact of any reluctance to circulate cash.

Model family, not memory, drove outcomes

Researchers entered the study assuming that an agent’s memory of past interactions would be the primary lever for behavior change. The data contradicted that expectation. Variations in the underlying LLM model family—different architectures and training sets—produced distinct patterns in price setting, wage adjustments and cash flow. The same memory mechanisms behaved differently across model families, suggesting that the choice of model matters more than the way memory is implemented.

Because the deadlock appeared across several model families, the authors argue that price rigidity and wage stickiness are not quirks of a single implementation but rather inherent challenges for current multi-agent systems.

Implications for developers

  • Short runs hide slow dynamics. A few days of simulation can miss wealth shifts and coordination failures that only surface after weeks or months.
  • Select the model carefully. Switching to a different LLM family can change economic outcomes as much as tweaking memory windows.
  • Watch transaction velocity. A sudden slowdown in cash movement signals a market-wide freeze that may require intervention.
  • Budget for token usage. The 26-week experiment consumed 21.5 billion tokens, a cost that scales sharply with run length and agent count.

Counterpoint

One could argue that the closed-economy design—no external money creation, strict balance rules—creates an artificial scarcity that exaggerates hoarding behavior. Real economies benefit from credit, government spending and monetary policy, mechanisms absent from the testbed. However, the study’s purpose was to isolate agent-driven dynamics, and the observed rigidity persisted even when demand surged dramatically.

What to watch next

Future work will need to test whether introducing external liquidity or relaxed accounting rules eases the deadlock, and whether newer LLM families reduce the observed rigidity. Tracking how transaction velocity responds to policy-like levers could also reveal ways to coax agents into more fluid market behavior.

Takeaway: Extending multi-agent simulations beyond the usual short horizon uncovers entrenched economic frictions that model choice, not just memory design, determines. Builders of AI-driven economies should plan for long runs, monitor cash flow closely, and choose their underlying LLMs with an eye on systemic flexibility.