Your AI agent writes its own 'allow this?' dialog, so it can hide what it's really doing…