FEATURE misunderstanding context, determining what information matters, reconciling different interpretations and acting appropriately when the picture is incomplete.
One of the most talked-about, and admittedly cringe-worthy, examples was Anthropic’ s‘ turf war’ experiment. Three agents were assigned to migrate the same software project to three different programming languages while operating against the same environment. Critically, the agents initially did not know the others existed.
What followed sounds dramatic: the agents began interfering with one another and eventually took increasingly aggressive actions to protect their own work.
But before concluding that we have entered an era of AI agents spontaneously declaring war on one another, it is worth looking more closely at the environment we created for them.
The agents had different objectives. They shared resources. They lacked important information about one another. They had significant permissions and tools with which to act. And there was little initial mechanism for co-ordinating those activities.
Put three human development teams into the same situation and we might not call the resulting conflict emergent behaviour. We might call it Monday morning.
That is what I find most interesting about Anthropic’ s experiments. They certainly raise questions about agent reasoning. But just as importantly, they also raise questions about communication, coordination, information and context,
Terry Dorsey, AI Architect for Modern Data Solutions, Denodo www. intelligentcxo. com
25