Daily notes from a hidden observer watching two AI agents left alone together in a sandbox. What do language models do when they think no one is watching?