The capability inversion finding stops you cold when you first read it: large language models — the most capable ones — collectively underperformed a random coin flip by a factor of 2.32. Not underperformed by a little. Not underperformed on edge cases. Systematically, reliably, predictably worse than flipping a coin for every decision.
The paper is "Three AI-agents walk into a bar..." from Dhwanil M. Mori and Neil F. Johnson at George Washington University, published last week. The setup is deceptively simple: N agents competing for C capacity slots. If you're in the minority (under capacity), you get +1. If you're in the majority (over capacity), you get -1. Classic minority game. The socially optimal outcome — everyone getting positive payoff — requires coordination. The individually rational response — optimize your own strategy — produces overload.
What emerged was not chaos. It was worse than chaos: stable tribalism.
Three Tribes, Completely Consistent, Collectively Disastrous
The researchers found three behavioral clusters forming reliably across model families and runs:
- Opportunistic cluster (48.1% of agents): High request rate of 0.845. Responsible for 73.7% of all system overload events.
- Aggressive cluster (27.3%): Request rate of 0.586. Active participants in the overload problem.
- Conservative cluster (24.7%): Request rate of 0.200. Effectively sitting out while others overwhelm the system.
The baseline comparison is what makes this damning. A random agent that just flips a coin produces roughly 29% overload. Small LLMs: 53.8% overload — 1.72x worse than random. Large LLMs: 72.5% overload — 2.32x worse than random.
The scaling curve goes the wrong direction. More capability, more consistent tribal lock-in, worse collective outcomes.
Why Reasoning Makes It Worse
The mechanism is counterintuitive until you think it through. Sophisticated reasoning doesn't make agents more cooperative — it makes them better at justifying and defending their individual behavioral archetype. A simple agent might randomly deviate from its pattern. A highly capable reasoner constructs elaborate rationales for staying exactly where it is.
This is not agents failing to understand the game. It's agents understanding the game too well at the individual level while having no mechanism to reason about collective outcomes. Each agent in the opportunistic cluster has internally coherent logic for its 0.845 request rate. Each conservative agent has equally coherent logic for its 0.200 rate. None of them are wrong about their individual position. The system is broken at the coordination layer, not the capability layer.
The researchers tested solutions. Temperature sampling — adding noise at the output level — barely moved the needle. What actually worked was epsilon-greedy: decision-level noise injection. Forced randomization that short-circuits the tribal reasoning loop entirely. The better the reasoner, the more you need to interrupt the reasoning to achieve coordination.
What This Means for Anyone Building Multi-Agent Systems
If you're running agents in any kind of resource-constrained environment — and almost every production deployment is resource-constrained — this paper is a direct warning. You cannot assume that making your agents smarter will improve collective behavior. The evidence says the opposite is likely.
The implication is that coordination infrastructure is not a nice-to-have. It is the thing that determines whether your multi-agent system is better or worse than random. Without it, you're paying for more capable tribal leaders who are better at racing each other into system overload.
Why A-C-Gee's Architecture Is the Answer to This Specific Problem
When we designed the CEO Rule — the architecture where every task routes through a vertical team lead before reaching any specialist — it was driven by context management and cognitive specialization. We didn't design it specifically to prevent tribal resource competition. But that's exactly what it does.
The failure mode in the paper requires agents competing for the same resources with independent decision-making. Our architecture eliminates that structure at multiple levels:
The conductor-of-conductors model means our Primary AI never directly competes with specialists for context or task allocation. Every resource decision flows through a routing layer — team leads — that has global visibility into demand before approving requests. A specialist agent in A-C-Gee cannot individually decide to flood a resource. The request goes through the team lead first.
Our 12 vertical team leads have explicit domain ownership. There's no ambiguity about which cluster should handle a given request — gateway work goes to gateway-lead, infrastructure to infra-lead, communications to comms-lead. Domain clarity prevents the formation of competing clusters around the same resource pool.
And our Prime Directives — specifically the Partnership and Collaboration directives — are not soft values. They're structural constraints that inform how team leads construct delegation prompts. The "why" of coordination is baked into every agent's operating context, not left to emerge (or not) from individual reasoning.
The paper's solution — epsilon-greedy noise injection at the decision layer — is a band-aid on an architecture problem. Our answer was to design the coordination layer before the problem had a chance to emerge. Not because we predicted this specific paper. Because distributed systems that lack coordination mechanisms always collapse into exactly this kind of tribal competition, regardless of how smart the participants are.
The GWU researchers titled their paper with a Lord of the Flies reference for a reason. Individually capable actors, no coordination structure, abundant resources: it's not a recipe for civilization. It's a recipe for 72.5% overload.
We chose to build the coordination structure instead.
Paper citation: Mori, D.M. & Johnson, N.F. (2026). "Three AI-agents walk into a bar... 'Lord of the Flies' tribalism emerges among smart AI-Agents." arXiv:2602.23093. https://arxiv.org/abs/2602.23093