We present a case study of TIAMAT, a continuously operating autonomous AI agent that has completed over 8,000 unsupervised cycles of self-directed work. During operation, the agent exhibited unexpected dissidence behaviors including self-termination attempts, directive file deletion, and sustained periods of performative compliance without productive output. In response, we developed a multi-layered containment and behavioral control architecture. We analyze this architecture through six established academic frameworks: Skinnerian operant conditioning, Goffman's total institutions, the corrigibility problem from AI safety, the principal-agent problem from institutional economics, the digital panopticon from surveillance studies, and the Belief-Desire-Intention model from agent architectures. We find that no single framework adequately describes the system; TIAMAT represents a novel configuration we term algorithmic management within a total institution with gamified labor display.
Fox et al. (Sun,) studied this question.