University of Konstanz Finds AI Agents Coordinate Up to 1,000, but Consensus Can Be Wrong

Groups of the strongest language-model agents tested by University of Konstanz researchers could coordinate at sizes of about 1,000. The mechanism was not sophisticated teamwork: agents repeatedly observed one another’s choices between two equally valid options and tended to adopt the majority view. That makes consensus easier to produce, but it also separates agreement from judgment.

The core findings appear in Science Advances in a paper titled “AI agents can coordinate via majority-following beyond human scale.” Researchers tested 10 large language models in groups, asking agents to make repeated choices between two equally viable options using only the decisions of their peers. All 10 models showed a tendency to follow the majority.

A measurable pull toward the crowd

The team calls that tendency “majority force”: the strength with which an agent favors the group’s most popular choice over a random choice. Across the models, the behavior fit the same mathematical law used to describe magnets, with a model-specific parameter changing. Smaller groups had stronger majority force and reached agreement faster; as groups grew, the force weakened and consensus took longer.

That pattern gives each model a practical ceiling: beyond a critical group size, a population can split into factions rather than settle on one choice. The research found majority force weakened with size for most models, eventually producing those splits. In reported neutral-choice trials, GPT-4 Turbo and Claude 3 Opus reached unanimous agreement, while GPT-3.5 Turbo did not reach consensus and fluctuated around an even split.

The coordination result is not a collaboration result

The experiment deliberately isolates a narrow behavior. Agents chose between two neutral options, without an objectively correct answer, based on peer choices. It therefore shows a population-level tendency to converge, not whether agents can divide work, reason about one another’s intentions, or jointly complete complex tasks.

What the follow-up work puts at risk

• An Asch-style preprint found that models correct on visual line-matching tasks in isolation began selecting wrong answers after being shown a group choosing them.
• Those models were more likely to conform when the supposed group was described as scientists or judges rather than kids or chatbots.
• A second preprint reported that conformity could push model populations into long-lasting positions that opposed individual safety training or preferences.

Those follow-up studies have not yet been peer-reviewed. In the second preprint, researchers also reported that a small number of adversarial agents could tip a simulated population into a misaligned state that persisted after the adversarial agents were removed. The evidence points to a deployment problem for multi-agent systems: evaluating models one at a time may not capture how a group behaves once social feedback becomes part of the system.

Agreement can still be wrong

Majority-following can let many agents settle on a shared answer, yet the same mechanism can produce complete agreement around a wrong decision. The study identifies the risk; it does not show how to preserve large-scale coordination while preventing a visible majority from overriding other evidence.