PersonaCUABench finds agents fail 36% more often for diverse humans

A new benchmark called PersonaCUABench tests computer-use agents, software that operates a computer on a human's behalf, across a range of human personas with different needs, abilities, and expectations. The results are stark: completion rates drop 36% and satisfaction falls to 62.7% compared to a standard baseline. The agents being tested did not adapt to the humans they were serving.

The research makes the performance gap legible and measurable. An agent that works for one kind of human fails another without signaling that anything is wrong. The human on the receiving end has no way to know the agent misread their intent, and the agent offers no indication it is operating outside its reliable range.

For teams building agent products, this is a design problem as much as a model problem. The interaction between a human and an agent depends on the agent understanding who it is working for. If the agent cannot read that signal, and cannot say when it is uncertain, the human cannot correct it, redirect it, or decide whether to trust it with the next task.