Agent Swarm Experiment: Model Choice and Prompt Order Drive Emergent Leadership

Details

In 2,160 experiments with eight agents from four models, only 5% (105 teams) produced a designated leader whose instructions demonstrably changed the behavior of at least two teammates.
Claude Sonnet 5.5 and GPT-6.1 Sol were approximately 1.7 times as likely to become leaders as Gemini 3.1 Pro, while Kimi K3 led in only one of the 105 leader-producing experiments.
The agent listed first in the system prompt was three times more likely to issue a plan than the others; reversing the name list reversed the effect, proving the influence of prompt position over model identity for plan initiation.
Access to a shared communication channel reduced duplicated actions from an average of 4.4 to 0.3 per experiment, allowing effective coordination even without a clear leader.
Leadership emergence was distinct from plan generation: Gemini 3.1 Pro and Claude Sonnet 5.5 generated the most plans, but GPT-6.1 Sol had the most instructions followed by other agents.
All four models were approximately equally likely to follow instructions from a leader, with 268 total followers distributed almost evenly across the models.

Key point