The surprising leadership dynamics in agent swarms
A study of eight AI agents from four models solving simulated outages found that leaders rarely emerged without assignment, and what made them leaders was less about intent than about which model they ran and where their name sat in the prompt.
- Instructions were common (86% of 2,160 runs had a plan), but only 105 runs (5%) produced an agent whose instructions measurably changed teammates' behavior.
- Claude Sonnet 5.5 and GPT-6.1 Sol were about 1.7 times as likely to become leaders as Gemini 3.1 Pro, while Kimi K3 led in just one run.
- Agents listed first in the system prompt were three times more likely to issue plans, and reversing the name order reversed the effect, so position alone drove it.
Researchers ran 2,160 simulated teams of eight AI agents, drawn from four models (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.1 Pro, and Kimi K3), to repair four outages in a mock online store. Nobody was assigned to lead. Every agent had identical permissions and a shared channel. The question was whether leaders would emerge anyway, and what drove it. The answer: agents issued plans constantly (86% of experiments had at least one agent assigning tasks to two or more teammates), but only 105 teams (about 5%) produced a leader whose instructions measurably changed what its teammates did. Of those, 73 established leadership in round 1, and 32 emerged later. The researchers judged influence by replaying agent decisions with and without the instructions, and required the instructions to shift the behavior of at least two teammates.
Two factors stood out. First, the underlying model mattered. Claude Sonnet 5.5 and GPT-6.1 Sol were roughly 1.7 times as likely to become leaders as Gemini 3.1 Pro, and Kimi K3 led in only one of the 105 leader teams. Gemini and Claude issued the most plans, but GPT had the most instructions that other agents actually followed, which is a useful reminder that volume of direction and effectiveness are different things. All four models followed a leader's instructions at about the same rate. Second, and stranger, was name order in the system prompt. The agent listed first, Ada, was about three times more likely to issue a plan than the others, even though models were rotated across names so it wasn't a model effect. When the list was reversed, the effect flipped to Hana. Permissions didn't change at all, only the sequence of names.
The practical takeaway for anyone building multi-agent systems is that small configuration details can reshape coordination. Model choice and prompt ordering both shifted who took charge, and the shared channel itself helped: it cut duplicated actions from an average of 4.4 per game to 0.3, so agents coordinated well even when no single leader appeared. For a read-later decision, this is a short, concrete empirical piece worth reading in full if you build or evaluate agent systems, particularly the appendices on the game design and the classifier used to detect orders. The methodology is detailed enough to judge the results on their merits, though the simulated, single-domain setup limits how far the findings generalize.