Evolution beats internal deliberation for multi-agent AI constitutions, study finds
180 simulations show external optimization wins in collective action—but with a critical catch.
A new study from researchers Hershraj Niranjani, Ujwal Kumar, and Phan Xuan Tan, posted on arXiv in May 2026, systematically compares two approaches for creating behavioral constitutions in multi-agent AI systems: internal deliberation (agents self-govern through dialogue) and external evolution (rules optimized via fitness functions). They tested both methods across three social environments—a coordination grid-world, an iterated public goods game, and a bilateral trading market—running 180 total simulations. The results show that evolution significantly outperforms deliberation in collective-action scenarios (p < 0.01), but neither method improved outcomes in the trading market.
The study also revealed a critical nuance through multiplier ablation: when the pool multiplier was set to m=0.75, the evolved constitution forced value-destroying cooperation, making it the worst-performing method instead of the best. Another striking finding: across all thirty deliberation trials, agents never proposed punishment—a canonical cooperation-sustaining mechanism that evolution consistently discovered. This suggests external optimization excels at finding high-performance peaks, while internal self-governance trades peak performance for structural responsiveness to changing conditions.
- Evolution outperformed deliberation in collective-action settings (p < 0.01) across 180 simulation runs.
- At pool multiplier m=0.75, evolved constitutions backfired, becoming the worst method by forcing value-destroying cooperation.
- No deliberation run ever proposed punishment, while evolution reliably discovered it as a cooperation mechanism.
Why It Matters
This research guides how to design rules for autonomous AI agent swarms, balancing performance with adaptability.