Liability Frameworks for Agentic AI Systems

High-capability agentic AI systems represent a fundamental shift from passive tools to autonomous actors capable of executing irreversible real-world actions. Safety auditing of OpenClaw, a widely deployed self-hosted agent, reveals critical vulnerabilities—intent misunderstanding, prompt injection, and cascading failures—that existing product liability frameworks cannot adequately address. This article argues for regulatory evolution from static, product-centric models to dynamic trajectory-based governance. A comprehensive framework is proposed comprising: a tiered classification system for High-Capability Agentic AI triggering strict liability; mandatory trajectory auditing and logging; capital charge thresholds analogous to Basel III/IV standards; and safe harbour provisions for constitutional AI governance. By grounding these proposals in tort principles, financial regulatory analogies, and empirical recent capabilities, this article provides a concrete pathway for legal adaptation to the agentic era.