The crash that vanished: control and emergence in a five-model economy
A hackathon project found that a market crash disappeared when one model was replaced with five different models. The fix was to override the settlement price, not the agents.
Intelligence analysis by GPT-5.4 Mini
The author rebuilt a tiny agent economy with five different small models and discovered that a crash seen in the one-model version vanished. After several failed attempts to shock the market, the reliable fix was to author the crash at settlement, where later trading could not undo it.
It is like trying to make a toy town panic by shouting, but the new crowd just keeps their snacks. The builder learned that if a big moment must happen, it works better to change the rule at the end of the round, like flipping the scoreboard after play.
Analysis
Five models, one story that stopped working
The original version of the Wood simulation used one model to run five creatures, and a reskinned bank run caused one agent to dump honey, sending the price from 10 to 3. When the author rebuilt the system with five different small models from different labs, that crash no longer appeared.
Why the market behaved differently
The player role was also rebuilt: instead of a direct trader, the player became a shadow financier who could short a good, spread a true tip, and try to profit from the fall. In the new setup, though, the council of models interpreted the rumor differently. Rather than selling honey, they hoarded it. The result was scarcity, not a crash, and the short position lost money.
The author then tried three ways to force the outcome back: relying on rumor alone, flooding inventories with honey, and increasing the size of the short. None of them reproduced the original collapse in the live models. One key lesson was that a cheap rule-based simulator can flatter a bad fix; if it behaves differently from the real agents, the simulator is the one misleading the builder.
The seam that worked
The eventual solution was not to persuade the agents, but to move the panic into a deterministic seam after trading had cleared. The bank run now overwrites the reference price at settlement, so the price halves as an authored consequence rather than an uncertain behavior. That preserves the emergent layer above it, where the models still gossip, trade, and react, while making the critical outcome reliable.
The article’s broader point is that emergence is contingent, not permanent. A behavior seen in one population can disappear when the population changes. If a moment has to happen, the safest place to enforce it is downstream of the agents’ choices, not by pushing harder on their inputs.
Key points
- A crash seen in a one-model economy disappeared when the author switched to five different small models.
- The council of models reacted to rumors by hoarding honey, which pushed the price up instead of down.
- Three attempts to force the crash through upstream shocks all failed in live runs.
- The working fix was to author the crash at settlement, after trading had cleared.
- The article argues that emergence is useful, but reliable outcomes need a deterministic seam.
If this lesson holds, builders can keep rich agent behavior for flavor while still guaranteeing important outcomes at a controlled point in the loop. That could make agent economies easier to test and more dependable without removing their life-like behavior.
The downside is that impressive behavior from one setup may vanish when the model mix changes, so results may not generalize. It also shows that fast simulators can validate the wrong fix and waste live runs before the mismatch is discovered.
