They're just not very good at it.
They're good at extracting information and generating text, but not simulating logic gates.
They're also good at open ended tasks where the logic is not clear at the outset. But as agent skills are developed, the need for that open-ended decision making is a liability. We want the agents to reliably follow a particular execution pathway.
Perhaps the solution is a deterministic harness encoding an expert-informed logic which hand-holds the agent as it works.