security, identity, governance and the architecture surrounding the agents themselves.

Perhaps the most significant question isn ' t simply:“ Why did the agents do that?”

It is:“ Which capability actually failed or was never adequately established in the first place?”

That introduces a subtle but important distinction between what is commonly known and what is true within the context of a particular question. Information can be widely available, repeated and seemingly consistent without being sufficient to reach the right conclusion. Less visible information can fundamentally change that conclusion once the appropriate context is established.

And this question becomes even more interesting when we look beyond the turf war.

And this challenge extends well beyond multiagent experiments.

Anthropic’ s hidden-profile experiment presents a very different problem. Instead of agents competing over shared resources, agents were given different pieces of information. Some information was commonly available, while other facts were known only to individual agents. A critical part of the experiment was the commonly shared information could point the group toward the wrong conclusion. Getting to the right answer required the agents to discover and combine information that was distributed among them.

So basically, the information needed to reach the right answer was there. The challenge was determining what information mattered.

Generative AI has access to enormous amounts of broadly available knowledge. Enterprise AI introduces another information environment layered on top of it: internal data, documents, business definitions, policies, relationships, transactions and organisational context. What is generally known may not be what is relevant, or even correct, for a particular company, business process, customer, point-in-time or decision.

Suddenly, the problem is not simply access to information.

The system must determine what is relevant, what is sufficient, what should be trusted and

26 www. intelligentcxo. com