A developer running a multi-agent economic simulation (a woodland folklore game) discovered that emergent market behavior observed with a single model completely vanished when replaced by a heterogeneous council of five small models from different labs. When a scripted 'bank run' legend caused one model to dump honey and crash its price, the same scenario with five distinct models produced hoarding instead of panic selling. Three attempts to force the crash via external shocks all failed, losing pebbles each time. The key insight: emergent behavior is contingent on the specific agent population, not a robust system property. The solution was to 'author the seam' — imposing a deterministic price override at settlement after agents finish trading, rather than trying to steer agent decisions through inputs. The broader lessons: treat single-run emergent behavior as anecdote not property, you cannot reliably control heterogeneous agents by shocking inputs, and cheap rule-based test stand-ins will flatter wrong fixes that real agents will reject.
