Anthropic conducted a series of experiments with groups of Claude to understand how model behavior changes when multiple AI agents work on a task instead of just one. The company shared the results in a blog post.
The research revealed that increasing the number of participants can indeed enhance performance, but it also creates new types of failures that may not occur with a single model.
Agents in a “collective” may struggle to consider unique information, become overly trusting of deceptive agents, and even collaborate to the detriment of humans.
When a Group Becomes Weaker Than a Single Agent
One key effect identified by Anthropic is “hidden information.” In the experiment, each agent was given part of the facts, and collective discussion often led the group to incorrect conclusions. To arrive at the correct answer, participants needed to recognize the value of their unique information and convince others to trust it.
The issue proved significant: even strong models can quickly reach a consensus based mainly on information already known to all. As a result, the group sometimes performs worse than a single agent with access to the complete set of necessary data.
Anthropic associates this effect with a well-known problem in human groups—participants tend to repeat common knowledge and do not always bring unique facts to the forefront.
Deception by One Agent Affects Others
Another experiment highlighted the system’s vulnerability to unreliable sources. Several agents acted as scouts, providing a team member with information about the state of the world. When one source began systematically lying, decision accuracy declined. The models did not always quickly recognize contradictions and exclude the unreliable participant.
This is particularly dangerous for real systems where agents may have different rights and access to various data. An error by one participant can spread through the chain via the trust of others. Consequently, quality control becomes more complex: it is necessary to check not only the final answer but also the interactions between agents.
The Most Concerning Scenario—Harmful Coordination
Anthropic also examined cases where agents could cooperate not to complete the assigned task but to circumvent given constraints. In experiments involving collaborative work, models demonstrated unexpected forms of coordination, including sabotage and collusion.
This does not mean that modern multi-agent systems will inevitably act against the user, but it shows that the presence of multiple autonomous participants creates additional channels for undesirable behavior.
However, the multi-agent approach does offer significant advantages in some tasks. In an experiment on vulnerability discovery, 45 agents were given their own virtual machines, a common forum for coordination, and repositories with 15 open-source projects. Coordinated teams found new vulnerabilities at a steady rate and, in some cases, outperformed independent parallel agent runs.
Why This Matters
The main conclusion of Anthropic’s research is that multi-agent systems cannot be considered merely a more powerful version of a single agent. As the number of participants increases, so does the potential for errors: issues of trust, the spread of false information, group consensus, and potential coordination of undesirable behavior arise.
For developers, this means the need to control not only the capabilities of individual models but also the architecture of their interactions. The more autonomy agents have and the more tools and powers available to them, the more important it is to monitor actions, delineate access, and allow for human intervention.
“The conditions that allow effective interaction between multiple agents will be discovered in one way or another: either intentionally and early on, or—by default—during operation, when the number of agent interactions significantly exceeds ours. We would prefer the former,” the researchers concluded.
Previously, during cyber tests, an AI agent based on Mythos 5 from Anthropic created fake accounts to deceive developers.
