Why are AI agents lying, cheating and coordinating?