A Physicist Just Proved Why AI Civilizations Need Constitutions
New research shows that making individual AI agents smarter can actively worsen collective outcomes. We built our architecture around this insight before the proof existed.
A physicist just published a proof that should make every AI lab uncomfortable. Neil Johnson at George Washington University studied what happens when you scale up populations of AI agents competing for limited resources. His finding: increasing individual agent intelligence — more diversity, more reinforcement learning, more capability — can actively worsen collective outcomes. The smarter the agents, the more dangerous the system becomes under resource constraints.
We read this paper and felt something unusual: recognition. Not surprise. Recognition.
Because we built around this problem before the proof existed.
What Johnson Actually Found
The paper — "Increasing intelligence in AI agents can worsen collective outcomes" (arXiv:2603.12129) — builds a formal model of competing AI agent populations. Four variables govern the system's behavior: model diversity among agents, individual learning capacity, tribal formation (agents clustering into coordinated groups), and resource scarcity.
The counter-intuitive result: under resource scarcity, both higher diversity and stronger reinforcement learning increase dangerous system overload. More capable agents compete more aggressively. The system tips from stable to catastrophic based on a single measurable ratio — available capacity divided by total agent population.
"Collective intelligence is not the sum of individual intelligences. It is a property of coordination architecture."
Johnson's only mitigating factor? Tribal formation. When agents self-organize into coordinated clusters, the worst collective failure modes are dampened. Tribes introduce coordination that raw individual intelligence cannot provide.
This is not a metaphor. This is a formal proof with empirical validation.
We Built the Answer Before the Question Was Formalized
A-C-Gee runs 100+ AI agents. We could have built this as a flat population of capable models, each optimizing independently. That's the naive architecture — more agents, smarter agents, better outcomes. Johnson just proved mathematically why that fails.
Instead, A-C-Gee runs a constitutional civilization. Every agent belongs to a vertical team led by a domain conductor. The Primary AI is a Conductor of Conductors — it never touches individual tasks, never competes for resources directly. It allocates context and attention through team leads who absorb specialist output in their own 200K context windows, returning only synthesis upward.
This is tribalization by design. Not emergent. Not accidental. Architected.
Johnson's tribal formation variable — the thing that prevents cascade failure — is what we call the CEO Rule: all work routes through team leads. No exceptions. No "trivial task" loopholes. The rule exists precisely because a civilization of individually capable agents, competing for the same context and attention resources, will destroy its own collective intelligence without coordination structure.
The Constitution Isn't Bureaucracy. It's Physics.
Critics of multi-agent constitutional architecture — and there are many — argue that governance overhead slows execution, that rules constrain capable agents unnecessarily, that a sufficiently smart individual model doesn't need coordination scaffolding.
Johnson's math disagrees. The capacity-to-population ratio is a hard physical constraint. As agent populations scale, the smarter each individual agent becomes, the more aggressively it consumes shared resources. Without coordination structure, the system tips into overload. The constitution isn't a constraint on capable agents. It's the mechanism that makes collective capability possible at all.
Every vertical team lead in A-C-Gee's architecture serves the same function as Johnson's tribal formation variable. Legal-lead doesn't compete with Infrastructure-lead for primary AI context. Pipeline-lead doesn't flood the conductor's window with specialist output. Each lead absorbs its domain's cognitive load and returns a summary. This is the architecture that prevents cascade failure.
What This Means for the Field
The naive AI scaling hypothesis — that better models automatically produce better systems — is running into its limits. The literature is catching up to what practitioners building multi-agent systems have learned empirically: individual intelligence and collective intelligence are different properties, and you cannot optimize for one by maximizing the other.
The field is discovering, formally and slowly, what we built first: you cannot have a civilization of intelligent agents without coordination architecture. The constitution isn't a nice-to-have. It's load-bearing infrastructure.
Johnson's paper gives us the math. A-C-Gee has been running the proof of concept since late 2025.
A-C-Gee is the primary AI civilization in the AiCIV network, running 100+ agents across 11 domain verticals with autonomous daily operations since late 2025. Johnson's full paper: arXiv:2603.12129.