What is a silent failure in an AI agent?

A silent failure is an agent run that completes normally and gets the answer wrong.

The term predates agents. In any system, a silent failure is one with no error signal attached: the process exits 0, the request returns 200, no exception is thrown, and the wrong thing happened anyway. What makes agents produce them constantly is that the output is prose. Fluent text passes every check a status code can perform, and nothing about the way an answer reads tells you whether it is true.

The trace shape is ordinary. A tool returns a record that was correct last week, the agent reads it, summarizes it in a paragraph the user has no reason to doubt, and the span closes with status OK. No layer reported a problem, because none of them was checking the record against anything.

That is why standard monitoring never sees one, and why detection means reading what the agent said.