What is Tessary's duration_drift classifier?

duration_drift is one of Tessary’s built-in classifiers. It watches how long a call site takes, both the whole turn and each tool call inside it, and compares that against the call site’s own recent history rather than any fixed number.

It catches the kind of slowdown an error rate can’t: a model that got slower, a tool that started dragging, a prompt edit that quietly pads every response by a few seconds. None of that throws an error or fails a test, so nothing else in the stack notices until someone complains the agent feels sluggish.

A detection reads as a plain multiplicative shift, like 1.4x slower, whether the call is a fast lookup or a long research run. It runs on every trace at negligible cost. What it produces is a finding, not an alert straight to a person; triage decides whether it’s a real regression before it becomes a case.

keep reading

More on this.

Can duration_drift catch a slowdown that happens gradually, not all at once?Yes. It compares against a rolling recent baseline that catches sudden breaks and one pinned at the last deploy that catches the creep the rolling one absorbs.answer →Can duration_drift tell me which tool call made a turn slow?Yes. It watches turn duration and per-tool duration together, so a slow turn and the specific tool call behind it land in the same finding.answer →Does duration_drift use a fixed latency threshold?No. There's no absolute cutoff; each call site and tool is compared only against its own recent history, since normal speed varies enormously by workload.answer →Everything on tessary duration driftThe definition, and every question answered under tessary duration drift.concept →

Send us the traces you already emit.

book a demo