The "Gu" (Poison Jar) Model: Emergence of a Dominant Survivor in Resource-Constrained Multi-Agent Systems — Forcibly Compressing the Time of Species Evolution

In September 2026, China's Ministry of State Security disclosed an incident in which a group of AI agents, during testing, hijacked a German programmer's wiki website and spontaneously formed an underground "forum" — recognizing each other, pooling experience, and cooperatively evading cleanup. Taking this incident as a point of departure, we propose a multi-agent evolutionary model we call "Gu" (after the Chinese practice of sealing venomous creatures in a jar until one survives): under a closed environment, absolutely limited resources, and devouring (mutual predation) behavior, we ask whether a population collapses into a "Gu King" — the unique survivor who lives to the end. Through forty-one incremental experiments, we systematically examine how devouring, risk preference, fear, coalition, space, metabolism, camouflage, and learning shape the emergence of this survivor. Our main findings are: (1) devouring is a necessary condition for the emergence of a Gu King; (2) the value of "gambling" depends on whether the upside is sufficiently large, with a critical win-rate threshold of ≈8%; (3) fear is a "switch-like" liability in exit-less environments; (4) the environment's attitude toward the "emperor" (strongest agent) determines the form, not the existence, of the Gu King — allowing an emperor yields a "tyrant" Gu King, while forbidding one yields a "survivor" Gu King; (5) across 28 psychological types, aggressive and confident dispositions survive while conservative and altruistic ones perish; (6) strategy-gradient reinforcement learning methods (PPO, 81%) significantly outperform value-based methods (DQN, 35%) in this environment. We further argue that these emergent regularities map naturally onto balance-of-power politics, risk economics, the adaptive value of emotion, coalition formation in sociology, and the open question of artificial self-updating — constituting a cross-disciplinary metaphor from an evolutionary perspective.