Autonomy, Defect Coupling, and Governance in Large Language Model Agents

Abstract

LLM-based agents are increasingly connected with memory, external tools, authorization rules, and communication channels. This shift makes local model errors more consequential, because unsupported statements may become plans, tool calls, memory entries, or messages sent to other agents. This paper develops a conceptual framework for explaining how native defects in LLM-based agents become system-level risks. Its main contribution is to connect previously separate discussions of hallucination, trust, memory, communication topology, and governance into an evidence-status-aware defect-coupling framework. The framework organizes the reviewed literature around four coupling mechanisms: generation-propagation coupling, trust-authorization coupling, knowledge-memory coupling, and control-communication coupling. The analysis also distinguishes directly grounded metrics, author-proposed operationalizations, and cross-domain analogies from classical multi-agent system control. Based on this framework, the paper compares governance strategies such as self-verification, debate, validator agents, guardian agents, memory observability, dynamic topology, and human-in-the-loop escalation. The analysis suggests that effective governance should be adaptive, evidence-status-aware, and authority-preserving.