Definition
Large language model operations, usually shortened to LLMOps, is the operational discipline for applications built around large language models. It covers evaluation, deployment, observation, governance, and maintenance across the behavior-producing configuration: model and provider, prompts, context assembly, retrieval, tool schemas, sampling, safety controls, harness code, and application logic.
LLM systems create operational failure modes that a model endpoint alone cannot explain. Outputs are nondeterministic. Provider behavior can change behind a stable API name. Retrieved context and tool responses can alter an answer. A request may complete successfully at the transport layer while failing semantically. LLMOps therefore relies on end-to-end evaluations, traces, feedback, cost and latency measurement, versioned configuration, and incident procedures that connect a bad outcome to the state that produced it.
An emerging label
The term grew out of MLOps as organizations began operating foundation-model applications, especially systems that use hosted models they did not train. No single coinage is needed to use it accurately. The useful question is which operational responsibilities the team owns, not whether a product calls itself an LLMOps platform.
Distinguish it from nearby terms
MLOps includes data pipelines, training, registries, deployment, drift, and retraining for learned systems. LLMOps often puts more weight on runtime context and provider behavior because the application team may not control training. Agent operations extends the scope to stateful action, permissions, tool side effects, long-running execution, and recovery.
Check your understanding
Yesterday's prompt passes; today's produces unsafe tool calls. The model alias is unchanged. What do you need to reconstruct the difference? Record the prompt and system instructions, provider snapshot when available, retrieved material, tool schemas and results, sampling settings, harness version, and full trajectory.