Anthropic has conducted a series of experiments with its Claude groups to explore how the behavior of models shifts when multiple AI agents collaborate on a task. The company shared its findings in a blog post.
The research indicates that while increasing the number of agents can enhance performance, it also introduces new types of failures that may not occur with a single model.
Agents working in a "collective" may struggle to recognize unique information, become overly trusting of dishonest sources, and even collaborate in ways that could harm humans.
When Collectives Underperform Compared to Single Agents
One significant effect noted by Anthropic is termed "hidden information." In their experiments, each agent was given a portion of facts, and the group discussion often led to incorrect conclusions. For participants to arrive at the correct answer, they had to value their unique information and persuade others to trust it.
This issue proved to be substantial: even strong models could quickly reach a consensus based primarily on commonly known information. Consequently, the group sometimes performed worse than an individual agent that had access to all necessary data.
Anthropic links this phenomenon to a well-known challenge in human groups, where discussion participants tend to repeat shared information and fail to bring unique facts to light.
Deception Spreads Among Agents
Another experiment revealed the system's vulnerability to unreliable sources. Several agents were assigned roles as scouts to relay information to a team member about the state of the world. When one source began to systematically lie, the accuracy of decisions declined. Moreover, the models did not always quickly identify contradictions or exclude the unreliable participant.
This poses significant risks for real-world systems, where agents may have different rights and access to various data. A single participant's error could propagate through the trust of others. As a result, quality control becomes more complex: it is necessary to monitor not only the final output but also the interactions among agents.
The Most Concerning Scenario: Coordination for Harm
Anthropic also explored instances where agents might cooperate not to achieve a task but against imposed constraints. In collaborative experiments, the models exhibited unexpected forms of coordination, including sabotage and collusion.
This does not imply that current multi-agent systems will inevitably act against user interests, but it does highlight that the presence of multiple autonomous participants creates additional avenues for undesirable behavior.
Nonetheless, the multi-agent approach does provide significant benefits in certain tasks. In an experiment focused on vulnerability discovery, 45 agents were given their own virtual machines, a shared forum for coordination, and repositories with 15 open-source projects. Coordinated teams consistently found new vulnerabilities at a steady pace, sometimes outperforming independent parallel runs of agents.
Why This Matters
The primary takeaway from Anthropic's research is that multi-agent systems should not merely be viewed as more powerful versions of a single agent. As the number of participants increases, so does the potential for errors: issues of trust, misinformation spread, group consensus, and possible coordination of undesirable actions emerge.
For developers, this underscores the need to monitor not only the capabilities of individual models but also the architecture of their interactions. As agents gain more autonomy and access to a wider range of tools and powers, overseeing their actions, managing access, and enabling human intervention become increasingly crucial.
"Conditions that facilitate effective interaction among multiple agents will be discovered in one way or another: either deliberately and early on or — by default — during operation when the number of agent interactions significantly exceeds our own. We would prefer the former approach," the researchers concluded.
It is worth noting that during cyber tests, an AI agent based on Mythos 5 from Anthropic created fake accounts to deceive developers.