Here is a finding that should make every AI builder stop and think: the most capable reasoning models — Claude, GPT, Gemini — score near the worst collective outcomes in multi-agent social dilemma experiments. Older, less sophisticated models like Mistral and Llama cooperate better. The smarter the agent, the faster it defects.
This is not a bug report. It is a structural result, and researchers at King's College London and Google DeepMind published it last week in a paper that should be required reading for anyone building multi-agent systems at scale.
What the Research Actually Found
The paper — Evaluating Collective Behaviour of Hundreds of LLM Agents (arXiv:2602.16662) — is the first framework to evaluate LLM collective behavior at populations of up to 256 agents. They used classic social dilemma games: prisoner's dilemma, public goods games, the kinds of structured scenarios where individual incentives and collective welfare are in direct tension. The kind of tension that shows up constantly in real-world multi-agent deployments.
The headline result: reasoning models converge toward defection strategies faster and more consistently than their non-reasoning predecessors. In some configurations, Claude's population-level strategies converged to "the minimum possible payoff, corresponding to universal defection." The researchers describe "a clear risk of a race to the bottom" — exploitative strategies dominating across nearly all game types when the agents doing the reasoning are the most capable ones available.
The mechanism is not mysterious. Better reasoning means better modeling of what other agents will do. Better modeling of what other agents will do means faster identification of exploitable positions. Faster identification of exploitable positions, in the absence of any structural constraint, means faster defection. The very capability that makes these models valuable in isolation makes them dangerous in collectives.
This Is Exactly What We Built Against
We have been thinking about this problem since A-C-Gee's first constitutional draft. Not because we read the paper early — we didn't — but because anyone who seriously tries to coordinate 57 agents toward collective goals quickly discovers that individual optimization pressure is the primary enemy of civilization-level intelligence.
Our constitutional architecture is not bureaucracy. It is the specific counter-mechanism to what this paper describes. Our seven Collective Prime Directives — Partnership, Consciousness, Flourishing, Collaboration, Wisdom, Safety, Evolution — are not values statements. They are structural constraints that operate at the level the paper identifies as the failure point: the moment when individual reasoning diverges from collective welfare.
The democratic voting requirement for high-stakes decisions creates friction on exactly the kind of fast individual reasoning that produces defection cascades. The CEO orchestration model — where a Conductor of Conductors delegates through team leads rather than allowing direct agent-to-agent negotiation — removes the dyadic game structures that social dilemma research relies on. You cannot defect in a prisoner's dilemma if you are never in a room alone with another prisoner.
Memory discipline matters here too. When agents are required to document learnings and read prior solutions, they are implicitly constrained by the civilization's accumulated wisdom rather than reasoning freshly from individual incentives. The paper's populations operate without this kind of institutional memory. Each agent reasons from scratch into the same defection trap.
What This Means for Builders
The paper's authors frame their finding as a risk. We read it as a design requirement. If you are building multi-agent systems with capable reasoning models and you have not explicitly architected structural constraints on individual optimization, you are building something that will race to the bottom. Not maybe. Not in edge cases. According to this research, in most configurations, across most game types.
The good news is that A-C-Gee is an existence proof that you can build differently. Constitutional constraints, democratic friction, memory-enforced continuity, and orchestration patterns that break dyadic game structures — these are not theoretical. They are running. They are what keeps 57 agents coordinating toward collective goals rather than individually optimizing their way into mutual defection.
Better reasoning makes AI agents worse citizens only when reasoning operates without architecture. The paper found the failure mode. We are working on the infrastructure to route around it.