The crash that vanished: control and emergence in a five-model economy

A researcher who rebuilt a multi-agent market simulation found that a previously documented emergent price crash disappeared when the model population was changed, revealing that the behavior was contingent on a single model’s disposition rather than a durable property of the system [1]. The original simulation, styled as a woodland bank run, saw an agent dump an asset and crash its price from 10 to 3 [1]. The author then reconstructed the environment, replacing one model running five creatures with a council of five distinct small models from labs including OpenAI, NVIDIA, and OpenBMB [1]. When the same rumor of a vault emptying was introduced, the heterogeneous council did not sell. Instead, the agents hoarded honey, driving the price up on scarcity and causing the short position to lose money [1]. “In an agent economy the reference price is not a dial you turn. It is the residue of what the agents actually choose to trade,” the author wrote [1]. The original crash was real but fragile, vanishing when the population changed [1]. Three subsequent attempts to force the crash using external shocks failed. Leaving the legend as a pure rumor produced no selling. Flooding inventories with a windfall, a tactic that worked against a rule-based test policy, was ignored by the live models. Increasing the size of the short only deepened the loss, with runs recording losses of 15, 26, and 27 pebbles [1]. The author noted that the cheap stand-in simulator “flattered a wrong fix,” and that when the stand-in and real agents disagree, the stand-in is the one lying [1]. The eventual resolution abandoned efforts to steer agent behavior through inputs. The crash was instead authored at the settlement seam, where the reference price is overwritten after all trading clears for the turn. This deterministic override halved the price reliably and turned the gambit into a profit of 40 pebbles [1]. The author argued that reliable outcomes come from choosing the precise seam for a deterministic override while leaving everything upstream free, using emergence for texture and authored control for moments that must happen [1]. The episode echoes a broader pattern in complex systems, where outcomes can hinge on the composition of participants rather than on universal laws. The original bank run that inspired the simulation began with the Wall Street crash of 1929, which ushered in the Great Depression, a period when U.S. unemployment rose to 25 percent and thousands of banks failed [2]. The author’s finding that a market crash could evaporate when the cast of agents changed underscores a lesson applicable beyond simulations: a single impressive run is an anecdote, not a property, until it survives a different population [1].