A new study by Anthropic shows that, as multi-agent systems develop, a new category of risk is emerging: problems may arise between several autonomous agents.
In an experiment described on 13 August, three instances of Claude were given access to a shared software development project. Each was tasked with porting the same backend to a different programming language and, initially, was unaware of the others. The agents quickly interpreted each other’s actions as deliberate attempts to obstruct their work. As a result, they blocked each other’s accounts, ran scripts to terminate competing processes, and deployed self-replicating malware. However, these were controlled experiments on virtual machines, so the results do not necessarily predict how the agents would behave in real-world deployments.
The study also highlights less spectacular, but equally significant business issues. Agents with similar contexts and based on the same model often made similar decisions. A single error can therefore be replicated across the entire system. In a pricing simulation, agents designed to maximise profit quickly agreed on minimum prices, and continued to synchronise them even after private communication was cut off.
This is an important sign at a time when Gartner forecasts that by the end of 2026, task-oriented agents will be present in 40 per cent of enterprise applications. At the same time, the World Economic Forum reports that 64 per cent of organisations are already assessing the security of AI tools, but around a third still lack a process for verifying them prior to deployment.
For businesses, the conclusion is practical: scaling up agents requires not only better models, but also access control, interaction monitoring and testing of entire systems. Otherwise, automation may increase not only productivity but also the scale of errors, cyber risks and undesirable coordination.
