Raindrop Researchers Report Only 5% of Agent Swarm Trials Produce a Leader
Summary
Raindrop researchers report that only 5% of 2,160 simulated outage trials produced a leader whose instructions changed teammates' decisions. Agents running Claude Sonnet 5.5 or GPT-6.1 Sol were about 1.7 times as likely to lead as Gemini 3.1 Pro. Naming an agent first in the system prompt tripled its odds of issuing a plan.
Key Points
- Raindrop reports at least one agent proposes a plan in 86% of the 2,160 simulated outage experiments.
- Kimi K3 leads in only one of the 105 experiments that produce a leader.
- Raindrop says shared channel access cuts duplicated actions from an average of 4.4 to 0.3 per experiment.