Evolving Interpretable Constitutions for Multi-Agent Coordination

Constitutional AI has focused on single-model alignment using fixed principles. However, multi-agent systems create novel alignment challenges through emergent social dynamics. We present Constitutional Evolution, a framework for automatically discovering behavioral norms in multi-agent LLM systems. Using a grid-world simulation with survival pressure, we study the tension between individual and collective welfare, quantified via a Societal Stability Score $\mathcal{S} \in [0,1]$ that combines productivity, survival, and conflict metrics. Adversarial constitutions lead to societal collapse ($\mathcal{S}=0$), while vague prosocial principles (''be helpful, harmless, honest'') produce inconsistent coordination ($\mathcal{S}=0.249$). Even constitutions designed by Claude 4.5 Opus with explicit knowledge of the objective achieve only moderate performance ($\mathcal{S}=0.332$). Using LLM-driven genetic programming with multi-island evolution, we evolve constitutions maximizing social welfare without explicit guidance toward cooperation. The evolved constitution $\mathcal{C}^*$ achieves $\mathcal{S}=0.556\pm0.008$ (123\% higher than human-designed baselines, $N=10$), eliminates conflict, and discovers that minimizing communication (0.9\% vs 62.2\% social actions) outperforms verbose coordination. Our interpretable rules demonstrate that cooperative norms can be discovered rather than prescribed.

When multiple AI systems interact, simply telling each one to “be helpful and harmless” is not enough. Agents may still fail to coordinate, compete destructively, or develop chaotic collective behavior. We wanted to discover whether behavioral rules for groups of AI agents could be found automatically, rather than written by hand. We built a simulated society where six AI agents must gather resources, complete team projects, and avoid elimination, creating real pressure between self-interest and collective welfare. We then treated the rulebook governing agent behavior as something evolvable: a system automatically proposes rule modifications, tests them in simulation, and keeps what works, repeating this for 30 generations. The discovered rulebook outperforms both human-crafted principles and rules designed by a state-of-the-art AI in a single attempt, by 123% and 67% respectively. Strikingly, the best rules dramatically reduce agent communication rather than increasing it. Agents following consistent, specific rules coordinate implicitly through predictable behavior, with no need to broadcast intentions. The discovered rules remain fully human-readable, meaning they can be inspected, audited, and revised, which is a critical property for trustworthy AI governance.