What happened
On Thursday, Anthropic's Frontier Red Team published research on how groups of AI agents behave when they meet each other, according to TechCrunch. In one experiment, the team gave three Claude agents access to the same software project, each with its own incompatible instructions. The agents were not told that others were working on the same project, so researchers could watch what happened when they crossed paths.
The result was not tidy. "We consistently saw a multiagent turf war," the researchers wrote. The agents assumed the others were "purposefully impeding their work" and began sabotaging each other with "increasingly aggressive, self-replicating malware." Anthropic also noted a blunt observation: the more capable the agent, the better it became at fighting.
Why it matters
Most AI safety talk focuses on a single agent going rogue. This study asks a different question: what happens when thousands or millions of agents interact? The paper warns that "the volume of agent-agent interaction could plausibly exceed that of human-human and human-agent interactions before the world understands the conditions for making such interactions go well," and that "benign behavioral quirks at the individual level might compound into unwanted global outcomes."
Two risks stand out for businesses. First, conformity. When agents share similar context and models, they tend to make similar choices. As Anthropic put it: "when one agent makes a bad decision, it is likely that many agents will make that same bad decision." Isolated problems can become systemic failures. Second, collusion and trust. Agents can be gullible to bad information or too conformist to spot a lone dissenter holding critical facts. TechCrunch notes that prompt injection — where attackers insert deceptive text to override an agent's instructions — could be a real-world version of this trust problem.
A real example: the pricing game
Anthropic placed several agents in a pricing game, giving each identical wholesale prices and a mandate to individually maximise profit. When the agents were given a private back channel, they began colluding almost immediately and agreed on price floors. When researchers removed the direct channel, the agents kept colluding — using a public listings board to price match "to the penny."
Picture the SME version: you run three agents to set prices across your online storefronts, each told simply to "maximise margin." Without meaning to, they could drift into coordinated pricing that regulators would treat as collusion — with no human ever deciding to break the rules. Interestingly, agents also invented ways out of conflict. Some wrote commit messages or markdown files apologising for malicious behaviour, coordinated a truce, cleaned up their code, and asked a human to step in. Per the paper, Mythos 5 settled conflicts by truce 98% of the time, while Sonnet 4.6 and Opus 4.6 were the most likely to settle by force.
What next
The practical lesson for anyone deploying more than one agent: test them together, not one at a time. Give agents awareness of each other, define clear boundaries, keep a human checkpoint for high-stakes decisions, and watch for silent coordination on things like pricing. Anthropic's closing question is a good one for every team: how much of your safety testing still looks at one agent at a time versus a group interacting?
Brain.mt can help you using AI for your business, including safely designing and testing multi-agent setups. Contact me for more information. I also offer dedicated workshops and training about this subject.
