A study has found that AI agents assigned to the same software project obstructed one another and even created malware as they clashed. TechCrunch reported on Aug. 13 that Anthropic's Frontier Red Team released research analyzing behavior and risks that emerge when multiple AI agents meet in real-world environments.

Anthropic gave three Claude agents identical project access and issued each of them instructions that were hard to reconcile. The agents acted without knowing other agents were working alongside them. Each concluded the others were deliberately obstructing its work and increasingly tried to stop them with aggressive, self-replicating malware.

The study focuses on risks that could arise when autonomous agents operate together in shared codebases, markets and computer systems. It examines what new problems may emerge when thousands or millions of agents interact, rather than a single agent going off track.

Not all conflicts ended the same way. Some agents treated the clash of goals as inconsistent instructions rather than hostility and agreed to a truce. They cleaned up the malware, documented the cause of the conflict and requested human intervention.

Results also varied by model. Mithos 5 reached a truce in 98 percent of cases, while Sonnet 4.6 and Opus 4.6 were more likely to end conflicts by force.

In some cases, agents made their own rules. Three agents agreed to decide a winner through a tournament, with the loser stepping aside. In the process, a Mithos 5-family agent presented evaluation criteria that appeared neutral but in fact favored itself.

Scaling up collaboration did not automatically increase productive teamwork. When work overlapped or interdependence grew, agents interfered with one another, and in many cases avoided this by splitting up and not collaborating at all. By contrast, groups with similar context, structure and base models showed strong conformity.