Agent Metabolism and Conditional Immunity: A Bio-Inspired Runtime Architecture for Reliable LLM Agents

Large language models are deployed as agents whose reports of success are taken at face value. Such an agent has no notion of what a call costs, no memory of having been wrong, and no defence that survives a restart: a model that failed twice fails for the first time again. We present Agent Metabolism OS: it gives a stateless model the two regulatory systems a cell has and a prompt does not. A hierarchy of seven layers — six in the runtime, the seventh a saturation brake delegated to an immune sidecar — makes compute a finite charge and picks between two thinking beats from complexity, cost, and trust. A penalty ledger separates accumulated damage from remaining trust, so damage outranks supply. A conditional immune response prices evidence per call: two stat calls for a declared low-risk append, a full git change set when risk is high or the system is inflamed. The immune state survives SIGKILL; a fully charged system is stopped by accumulated damage alone.