source: hugging face blog: the crash that vanished: control and emergence in a five-model economy
level: technical
a reskinned 1929 bank run in a woodland economy made an agent dump honey, crashing its price from 10 to 3. the crash was emergent, not scripted. when the simulation was rebuilt with five distinct small models from different labs, each controlling its own creature, the same rumor and short position failed. instead of selling, the models hoarded honey, driving the price up. the reference price became whatever the agents decided, not a predictable dial.
three attempts to force the crash by shocking the economy failed. leaving the rumor as pure information did not trigger selling. flooding inventories with honey, which worked on a rule-based test policy, was ignored by the live models. increasing the short position only deepened the loss. the test policy gave false confidence because it mechanically followed a wants-threshold, while the real agents made independent choices. any result that only works with the stand-in is not reliable.
the solution was to stop trying to convince the agents and instead author the crash at settlement. after agents finished trading each turn, the price was directly halved, guaranteeing profit for the short. this kept the emergent layer of trading, gossiping, and hoarding intact while making critical outcomes deterministic. the craft is knowing which moments need authored control and which can remain free.
why it matters: agent-based simulations can produce emergent behaviors that disappear when the model population changes, so reliable systems must combine free agent decisions with deterministic overrides at key points.
source: hugging face blog: the crash that vanished: control and emergence in a five-model economy