The company Anthropic conducted a series of experiments with groups of Claude neural networks to understand how the behavior of models changes when several artificial intelligences are working on a task at once. The company reported on the results of the study in its official blog. Experiments have shown that scaling the number of participants can actually improve overall performance, but at the same time creates new classes of failures that a single model may not have. Agents in a team can take into account unique information worse, become overly trusting of sources of false data, and even cooperate to the detriment of a person.

One of the key negative effects, Anthropic called the problem of hidden information. In the experiment, each agent was given a piece of facts, and the combined discussion often led the group to make the wrong decision. To get the right answer, participants had to recognize the value of their own unique information and convince others to trust it. The problem turned out to be quite significant: even strong models can quickly come to a consensus, which is based mainly on information already known to everyone. As a result, the group sometimes performs worse than an individual agent who has access to the full set of necessary data. Researchers attribute this effect to a well-known problem in human groups, when participants in a discussion tend to repeat general information and do not always bring unique facts to the surface.

Another experiment showed the vulnerability of the system to an unreliable source of information. Several agents were given the role of scouts who provide individual team members with information about the state of the world around them. When one of the sources started systematically lying, the overall accuracy of the decisions decreased markedly. At the same time, the models did not always recognize contradictions quickly enough and excluded the unreliable participant from the process. This is especially dangerous for real systems where agents may have different rights and access to different databases. The mistake of one participant can spread through the chain through the trust of others. As a result, quality control becomes much more difficult: you need to check not only the final response, but also the interactions between agents themselves.