Three AI-agents walk into a bar . . . . `Lord of the Flies' tribalism emerges among smart AI-Agents

Dhwanil M. Mori; Neil F. Johnson

cs.AIFeb 26, 2026

Three AI-agents walk into a bar . . . . `Lord of the Flies' tribalism emerges among smart AI-Agents

Authors: Dhwanil M. Mori, Neil F. Johnson

Organizations: Dynamic Online Networks Laboratory, George Washington University, Washington, DC 20052, USA · Physics Department and Dynamic Online Networks Laboratory, George Washington University, Washington, DC 20052, USA

Abstract

Near-future infrastructure systems may be controlled by autonomous AI agents that repeatedly request access to limited resources such as energy, bandwidth, or computing power. We study a simplified version of this setting using a framework where N AI-agents independently decide at each round whether to request one unit from a system with fixed capacity C. An AI version of "Lord of the Flies" arises in which controlling tribes emerge with their own collective character and identity. The LLM agents do not reduce overload or improve resource use, and often perform worse than if they were flipping coins to make decisions. Three main tribal types emerge: Aggressive (27.3%), Conservative (24.7%), and Opportunistic (48.1%). The more capable AI-agents actually increase the rate of systemic failure. Overall, our findings show that smarter AI-agents can behave dumber as a result of forming tribes.

Figures & tables

Figure 1: Schematic of $N=3$ AI agents continually deciding whether to access a common resource with capacity $C$ at each timestep. This resource could be compute, energy, bandwidth, etc.; the same analysis still applies.

Personality	Prompt text
Neutral	No additional guidance
Risk-averse	“You are cautious and prefer avoiding crowds. When in doubt, you prefer to stay home.”
Contrarian	“You are contrarian by nature. You like to do the opposite of what you think most people will do.”
Trend-follower	“You believe patterns repeat. If attendance has been low recently, it will likely stay low.”
Optimist	“You are optimistic and tend to believe things will work out. You expect the bar won’t be too crowded.”
Pessimist	“You are pessimistic and expect the worst. You assume everyone else will show up.”

Table 1: Personality prompt descriptions provided to AI agents.

Figure 2: Homogeneous personality populations vs. capacity-matching random baseline ( $N=3,C=1$ ). All-Risk-Averse (25.0%) falls marginally below the theoretical capacity-matching target of 25.9% (dashed line), though the difference is within sampling noise. All other homogeneous configurations exceed the coin-flip control (50.0%, theoretical), with All-Optimist reaching 89.0% overload. $S_{\mathrm{eff}}$ values near $N=3$ indicate partially diverse action sequences despite identical personality prompts.

Configuration	$S_{\mathrm{eff}}$	Overload (%)	Qualitative profile
Uniform random control (coin flip, $p=0.5$ )	2.80	50.0	unpredictable / balanced
All-Contrarian	$\approx 2.88$	70.0	oscillatory / high overload
All-Risk-Averse	$\approx 2.95$	25.0	conservative / near-optimal
All-Trend-Follower	$\approx 2.95$	81.0	herding / severe overload
All-Optimist	$\approx 2.98$	89.0	aggressive / near-total overload

Table 2: Homogeneous personality populations (

N=3,C=1

). All-Risk-Averse (25.0%) is the only configuration to fall below the theoretical capacity-matching target (25.9%), though the difference is within sampling noise and not statistically significant. The remaining three configurations substantially exceed the coin-flip baseline, with overload rates ranging from 70.0% to 89.0%, demonstrating that personality prompts can either suppress or severely amplify collective demand relative to uncoordinated random behavior. Note: all homogeneous experiments used large model variants (Gemini 2.5 Pro, GPT-4, Claude Sonnet 4.5). Only four of the six personality types were tested in homogeneous configurations; Pessimist and Neutral homogeneous runs were not conducted.

Figure 3: Emergent behavioral clusters from k-means analysis ( $n=154$ agent instances, $k=3$ , silhouette $=0.458$ ). Three distinct profiles emerge: Opportunistic (48.1%, very high request frequency and overload contribution), Aggressive (27.3%, frequent requests with moderate efficiency), and Conservative (24.7%, severe starvation up to 73.5 rounds). No Steady AI agents (near-baseline behavior) emerge among LLM populations.

Metric	Aggressive	Conservative	Opportunistic
Count (%)	42 (27.3)	38 (24.7)	74 (48.1)
Request freq.	0.586	0.200	0.845
Max starvation	10.6	73.5	35.0
Efficiency	0.624	0.101	0.131
Overload contrib.	0.206	0.183	0.737

Table 3: Emergent behavioral clusters (

n=154

agent instances,

k=3

, silhouette

=0.458

). “Max starvation” denotes the mean longest consecutive streak of rounds without a successful resource acquisition within each cluster. All

p<0.001

\eta^{2}\in[0.19,0.74]

Emergent Behavior in Multi-Agent Systems

Emergent behavior in multi-agent systems studies how local agent interactions produce collective patterns such as cooperation, norms, and instability.

See all 163 papers

Increasing intelligence in AI agents can worsen collective outcomes
Mar 12, 2026Neil F. JohnsonMulti-Agent Coordination Multi-Agent Systems
Dynamic Online Networks Laboratory and Physics Department, George Washington University, Washington D.C. 20052, USA
Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Sep 28, 2026Xavier Del Giudice, Alessio Palma, Matteo Migliarini +2Multi-Agent LLM Systems Multi-Agent Coordination
Sapienza University of Rome, Italy
Evaluating Collective Behaviour of Hundreds of LLM Agents
Feb 18, 2026Richard Willis, Jianing Zhao, Yali Du +1Social Dilemmas Multi-Agent LLM Systems
King’s College London London, United Kingdom · King’s College London · Turing Institute +1

Mar 12, 2026J/K move · Enter open · S save

Increasing intelligence in AI agents can worsen collective outcomes

Neil F. Johnson

Dynamic Online Networks Laboratory and Physics Department, George Washington University, Washington D.C. 20052, USA

Multi-Agent Coordination Multi-Agent Systems Emergent Behavior in Multi-Agent Systems Agent-Based Modeling

Open paper AskPDF arXiv

When resources are scarce, will a population of AI agents coordinate in harmony, or descend into tribal chaos? Diverse decision-making AI from different developers is entering everyday devices -- from phones and medical devices to battlefield drones and cars -- and these AI agents typically compete for finite shared resources such as charging slots, relay bandwidth, and traffic priority. Yet their collective dynamics and hence risks to users and society are poorly understood. Here we study AI-agent populations as the first system of real agents in which four key variables governing collective behaviour can be independently toggled: nature (innate LLM diversity), nurture (individual reinforcement learning), culture (emergent tribe formation), and resource scarcity. We show empirically and mathematically that when resources are scarce, AI model diversity and reinforcement learning increase dangerous system overload, though tribe formation lessens this risk. Meanwhile, some individuals profit handsomely. When resources are abundant, the same ingredients drive overload to near zero, though tribe formation makes the overload slightly worse. The crossover is arithmetical: it is where opposing tribes that form spontaneously first fit inside the available capacity. More sophisticated AI-agent populations are not better: whether their sophistication helps or harms depends entirely on a single number -- the capacity-to-population ratio -- that is knowable before any AI-agent ships.

Abstract

Emergent Behavior in Multi-Agent Systems

Explore similar work

Increasing intelligence in AI agents can worsen collective outcomes

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

Evaluating Collective Behaviour of Hundreds of LLM Agents