Capability benchmarks may become less meaningful once agents operate across long execution horizons with external tools and permissions. The governance problem starts shifting toward execution boundaries and observability.