What is Tessary's duration_drift classifier?
It detects a call site taking longer than its own history says it should, covering both the whole turn and each tool call inside it, rather than measuring against any fixed number of seconds.
That’s the kind of slowdown an error rate can’t see: a model that got slower, a tool that started dragging, a prompt edit that quietly pads every response. Nothing fails, so nothing else notices until someone says the agent feels sluggish.
A detection reads as a plain multiplicative shift, 1.4x slower say, which is what makes it comparable whether the call is a fast lookup or a long research run. From there it’s a finding, and triage rules on it before it reaches a person.