Shared Structural Vulnerability in Agent-Only Interaction Systems

This paper formalizes a recurring structural vulnerability observed across contemporary agent-only interaction systems deployed in non-isolated socio-technical environments. Independent of implementation details, intent, or cultural framing, such systems consistently lack three critical governance elements: explicit human-in-the-loop control, defined stopping authority, and responsibility attribution. When these elements are absent, agent-to-agent interaction loops become self-reinforcing optimization structures that prioritize internal coherence and speed while structurally excluding verification, accountable intervention, and safe termination. Based on cross-context observational probing rather than internal system access, this work demonstrates that the vulnerability manifests as a governance failure rather than a technical malfunction or ethical disagreement. The risk is not speculative or future-oriented; it is operational and observable in present-day deployments where agent outputs propagate into broader AI systems and human discourse. The paper further situates this system-level failure mode in relation to inference-time governance failures previously formalized as the False-Correction Loop (FCL), and argues that explicit stopping authority and responsibility attribution are foundational requirements for safe agent-based system design. This work is intended for researchers and practitioners in AI governance, agent framework design, safety engineering, and socio-technical systems, and emphasizes design-level control architectures over intent- or narrative-based approaches.