Agent skills look great in benchmarks but fall apart under realistic conditions, researchers find