Agent Metabolism and Conditional Immunity: A Bio-Inspired Runtime Architecture for Reliable LLM Agents
Year: 2026
Type: preprint
Abstract: 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.
Author: liao, zhizhi(廖至智)
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225120860