The proliferation of managed LLM services has introduced possibilities for generative AI applications, but also increases the need for comprehensive observability and debugging capabilities. Chain-based architectures and prompt engineering techniques are being used to build LLM applications, however introducing non-deterministic LLM services into an application can be challenging due to the complexity of tracing chain execution steps. To overcome these challenges, using a span hierarchy and instrumentation tools such as Datadog's LLM Observability can provide visibility into the input and output of each chain step, enabling teams to track code and request errors, pinpoint root causes of unexpected responses, identify hallucinations or improperly structured input, and optimize LLM provider costs. By instrumenting chains with a span hierarchy and adding metadata and context, teams can quickly identify sources of errors and latency and ensure their chains are reliable and performant.