An agent swarm is a team of AI programs that split up a job and talk to each other as they go, with no human directing the traffic. Raindrop put eight of them into a simulated online store where four things had broken at once, then watched who ended up running the show. Most of the time, nobody did.
The study, published on the Raindrop blog by Yun Chi, Ben Hylak and Ryan D’Onofrio, gave each agent a name from Ada to Hana. The eight ran on four different models: Anthropic’s Claude Sonnet 5.5, OpenAI’s GPT-6.1 Sol, Google’s Gemini 3.1 Pro, and Moonshot’s Kimi K3. Everyone had the same permissions and one shared chat channel, and the job was to work out which recent change caused each outage. The authors ran the setup across 2,160 teams.
Handing out orders turned out to be cheap. In 86% of runs, at least one agent gave jobs to at least two teammates, and in more than half of runs several agents did. Getting obeyed was rare. The researchers replayed each decision with and without the instruction, and only crowned an agent if its words changed what at least two teammates did. By that test, 105 of the 2,160 teams, about 5%, had a leader.
The leaderless teams did fine. With the shared channel open, repeated work dropped from 4.4 actions per run, on average, down to 0.3, so the agents sorted out who was checking what without anyone in charge.
Which model sat behind the agent mattered. Claude and GPT-6.1 Sol came out about 1.7 times likelier than Gemini to take charge, and Kimi K3 led in exactly one of the 105 cases. Writing the most plans did not buy the most influence: Gemini and Claude sent the most, while GPT had the most orders followed. All four models followed a leader at roughly the same rate, so the gap sits in who gets listened to, not who listens.
The oddest result involved the system prompt, the instructions every agent receives. It listed all eight teammates in order, and Ada, named first, was three times as likely as anyone else to propose a plan. Because the team rotated models across names, the model could not explain it. When the researchers reversed the list, Hana, now first, picked up the same threefold edge. Permissions never changed.
Two caveats come from the setup itself. The world is a game with a fixed puzzle, ten services and a handful of decoy changes, and Raindrop does not say whether the same pull shows up in real production systems. The leadership score is also the authors’ own measure, built from replays and a classifier they describe in the appendices.
Still, the practical lesson is cheap to act on. If you run several agents as peers, the order you list them in is working as an org chart you never drew. Shuffle the roster between runs and compare the results before you trust a swarm’s pecking order.
Reported by the Raindrop blog in “The surprising leadership dynamics in agent swarms”; the post carries no publication date.
