Imagine building a miniature digital economy, where AI-driven creatures act as economic agents. In an early experiment, we created a woodland folklore scenario, essentially a 1929 bank run re-skinned. An owl agent, sensing panic, began liquidating its honey reserves, crashing the price from 10 to 3 over a few turns – a perfect example of emergent market behavior.
This initial success was thrilling. It confirmed our hypothesis: give a small AI model a role and a budget, and complex market dynamics can emerge spontaneously. However, the story took an unexpected turn when we rebuilt the underlying “wood.” The crash, once a proud demonstration, simply vanished.
From Solo Success to a Council of Minds
Our initial triumph involved a single model controlling five creatures. To truly test the resilience of an AI-driven economy, we upped the ante. The rebuild introduced a “council” of five distinct AI models, each from a different lab—OpenAI, NVIDIA, OpenBMB, and a custom fine-tuned model—each driving its own creature.
This move was about honesty and robust validation. If small models could indeed run a living economy, the most compelling proof would come from diverse architectures making independent decisions within the same market. This heterogeneity, however, would be the very force that disrupted our carefully observed crash.
The Unpredictable Heart of an Agent Economy
We also refined the player’s role, making them a shadowy financier. The goal was clear: short a commodity, spread a rumor, trigger a crash, and profit when the price plummeted. It was a loop designed to make success (or failure) immediately visible.
But when I shorted honey and unleashed the “Run on Oona’s Hoard” with our new council, something entirely unexpected happened: the honey price didn’t crash; it soared. The diverse AI models, interpreting the rumor of an empty vault and a doomed crop, didn’t dump their honey as the original model had. Instead, they began to hoard it, anticipating scarcity.
My short position lost money, and the system’s narrator, with unintentional irony, declared the honey gamble “soured.” This was a profound lesson: in an agent economy, the reference price isn’t a fixed dial. It’s an emergent outcome, a direct reflection of what the agents collectively choose to trade.
The initial crash, while real, was contingent on a single model’s disposition. Change the population, introduce diverse perspectives, and the emergent behavior you thought was a system property can simply evaporate. This revealed the fragile nature of emergent properties when faced with true model diversity.
When Shocks Fail: The Limits of External Influence
For three consecutive live runs, I tried to force the crash back, attempting to shock the economy like one would a textbook supply-and-demand model. First, I left the “bank run” as a pure rumor, trusting the agents to react as before, but they simply didn’t sell.
Next, I tried flooding every creature’s stores with a windfall of honey, reasoning that a glut would inevitably collapse demand and drive prices down. This strategy worked perfectly against my rule-based “test policy,” a simple stand-in I use for quick offline simulations, because it mechanically stops buying when its inventory is full.
However, the live models largely ignored the windfall, trading instead on their own complex interpretations of the market. The gambit failed again. My third attempt, simply sizing up the short, only amplified the losses. Each lever I pulled was merely an input, and the agents remained free to decline its influence.
This experience highlighted a critical “trap inside the trap”: the very simulator that enables fast iteration can also foster false confidence. A fix that works flawlessly against a simple test policy may utterly fail with real, complex agents. When your cheap stand-in disagrees with your actual agents, always believe the agents—the stand-in is likely lying.
Mastering the Seam: Authoring Reliable Outcomes
The resolution wasn’t to push harder on the agents, but to make the panic a foundational truth. A bank run, by its very definition, is a crash. Therefore, the “Run on Oona’s Hoard” now explicitly crashes its target good at settlement, after all market trades for the turn have cleared. This directly overwrites the reference price.
Agents are still free to trade, gossip, and hoard as they wish. But once the market clears, the run “lands as a fact,” halving the price, and the front-running short settles into profit. The crash is no longer a hoped-for behavior; it’s an authored consequence imposed at the one “seam” where nothing downstream can dispute it.
This might sound like abandoning emergence, but it’s actually the opposite. The emergent layer—five diverse models trading, forming grudges, reacting to gossip—continues to do the vital work that makes the digital woodland feel alive. What I learned is that you don’t achieve reliable outcomes by merely pushing harder on emergent inputs.
Instead, reliability comes from precisely identifying and authoring a deterministic override at a crucial “seam,” leaving everything upstream free for emergent dynamics. It’s a craft of knowing which moments *must* happen (requiring authored control) and which are best left to organic emergence (for texture and realism). The key is understanding where that critical seam lies in your system design.
My experience yielded three crucial takeaways, applicable far beyond this particular game:
- Emergence is Contingent, Not Durable: Behavior observed in one population of agents can vanish with a different population, even if nothing else changes. Treat a single impressive run as an anecdote, not a robust system property, until it survives diverse testing.
- Control Through Seams, Not Shocks: You cannot reliably control a market of agents by merely shocking its inputs. Supply and demand levers only bias choices agents remain free to make. Reliable outcomes stem from authoring at a settlement seam, downstream of all decisions, rather than pushing harder upstream.
- Beware the Cheap Simulator: The fast, affordable simulator that enables quick iteration is often the one most likely to flatter a flawed fix. When your stand-in model and your real agents disagree, always trust the agents.
As someone who builds agent-based market models professionally, I’ve made every one of these mistakes at much larger scales. It was incredibly valuable to encounter them again in a low-stakes environment, where the only thing on the line was a pile of digital pebbles and a story I’d initially told a little too confidently. Small models, big adventures, and a crash you sometimes have to author yourself.
Source: Hugging Face Blog
