AI agent skills look great in benchmarks but fall apart under realistic conditions according to researchers