Inherent and emergent liability issues in LLM-based agentic systems: a principal-agent perspective