Observability and Orchestration for Distributed LLM Agents Across the Cloud-Edge Continuum

This preprint presents a distributed-systems perspective on the deployment, observability, and orchestration of LLM-based agents across the cloud-edge continuum. It synthesizes recent work on agent systems, model serving, cluster scheduling, edge computing, distributed tracing, and zero-trust security. LLM-based agents increasingly combine model inference with retrieval, persistent state, tool calls, and interactions with other services or agents. When these components are distributed across devices, edge nodes, and cloud infrastructure, the application inherits familiar systems problems: placement, latency, state tracking, access control, failure recovery, and end-to-end observability. This paper examines those problems from a distributed-systems perspective. It brings together recent research on LLM agents and agent evaluation with established work on model serving, cluster scheduling, edge computing, distributed tracing, and zero-trust security. The resulting deployment view separates the agent runtime, state, tools, and compute resources while treating observability and policy enforcement as cross-cutting concerns. The paper does not introduce a new agent framework and does not report new performance measurements.