Raindrop ran 2,160 experiments with eight-agent swarms resolving simulated outages. Only 5% of runs produced a leader whose instructions changed teammates’ actions. Claude Sonnet 5.5 and GPT-6.1 Sol were 1.7 times likelier to lead than other models. Agents listed first in the system prompt were three times likelier to issue plans. Kimi K3 almost never led. All models followed instructions at roughly equal rates. Shared communication cut duplicated actions from 4.4 to 0.3 per experiment. Leadership emerged without a designated leader. The effect of name order reversed when the list was reversed. Developers should note that small configuration details can shape swarm behavior.
What changed
Agents in swarms were assumed to coordinate without emergent leadership or positional bias.
What it unlocks
Developers can influence agent leadership by selecting models and ordering names in prompts.
- 2,160 experiments
- 105 experiments with a leader
- 5% of experiments produced a leader
- 1.7x more likely for Claude Sonnet 5.5 and GPT-6.1 Sol
What you need to act on it
- a swarm of eight agents from four models
- a shared communication channel
- no designated leader
Sources