What does a behavior drift baseline catch, and what does it miss?

It catches a new or reordered step almost every time, and misses a real share of the times an agent quietly stops doing something it used to do.

The reason is structural, not a tuning problem. An agent calling a tool it’s never used, or taking a path it’s never taken, leaves something new in the trace for the baseline to notice. An agent dropping a step it always ran leaves nothing: the trace just looks like a shorter version of normal, and normal is exactly what a behavior baseline is built to let through. New actions, swapped steps, and reordered ones are the strong catches. A quiet omission is the classifier’s honest weak spot.

If a step matters for correctness and not just for shape, pairing this with a check that asserts the step happened covers the gap drift detection alone won’t.