LLM Observability platforms monitor LLM-powered applications — capturing prompts, completions, tool calls, agent traces, token counts, and costs — and provide evaluation, debugging, and quality-regression-detection tooling purpose-built for LLM behavior. Core capabilities include end-to-end trace capture across multi-step agent workflows, prompt and response logging with sensitive-data redaction, LLM-as-judge and metric-based evaluation frameworks, dataset management and regression testing, A/B comparison across prompts and models, cost and latency analysis per feature/model/customer, and user-feedback capture tied to traces. The category targets LLM-specific production concerns (hallucination, prompt injection, drift in qualitative output) that general-purpose APM does not address. Leading vendors include Langfuse, LangSmith (LangChain), Helicone, Braintrust, Traceloop, HoneyHive, and Galileo.