Why did my agent stop doing a step it always used to do?

Something changed upstream: a prompt edit, a tool going away, or a branch in the agent’s own logic taking a different path, and now it skips a step it used to run every time.

Tessary’s behavior_drift classifier is built to catch exactly this. It learns your agent’s normal action sequence per call site from its own history, and a dropped step is one of the departures it watches for, alongside a new step appearing or a call site trying a path it’s never taken. It doesn’t judge whether the resulting answer was still correct, only that the shape of the trace changed from what that call site normally does.

It’s also the hardest of the three to catch reliably, because a dropped step leaves nothing behind to notice; the trace just looks like a shorter, quieter version of normal. If the step matters for correctness, pair this with a check that asserts it happened, rather than relying on drift detection alone.