What Is AI Observability?

AI observability is the practice of recording what an AI application did on every request, in enough detail to explain any single answer and to see trends across many. For a large language model (LLM) application that means the prompt, the retrieved documents, each tool call and its result, the output, token counts, cost and latency, linked together as one trace. Classic monitoring tells you the service is up; observability tells you why the agent gave a wrong answer.

What Gets Recorded

A trace is a tree of spans. One user request becomes a root span, with child spans for each step:

  • Model calls. Model name and version, input and output messages, token counts, latency, finish reason.
  • Retrieval. The query, the documents returned, their scores.
  • Tool calls. The tool name, arguments, result and errors.
  • Agent steps. Which branch was taken, how many iterations ran, why the run stopped.
  • Feedback. User ratings, reviewer corrections, evaluation scores attached to the trace afterwards.

OpenTelemetry, the open standard for telemetry, maintains generative AI semantic conventions that define common names for spans, metrics and events from GenAI clients and from the Model Context Protocol (OpenTelemetry GenAI semantic conventions). Using them lets traces move between tools instead of being locked into one vendor’s format. Dedicated platforms such as LangSmith add LLM-specific views on top. Its documentation describes traces as the record of what agents did in production and the basis for debugging failures, monitoring quality and building evaluation datasets, and it works with many frameworks and model providers (LangSmith observability).

Signals to Watch

SignalWhat it shows
Tokens and cost per feature or userWhere spend comes from
Latency, including time to first tokenWhether the experience is acceptable
Tool error rateBroken integrations, bad arguments
Steps per agent runLoops and inefficient plans
Groundedness and feedback scoresWhether answers are supported and useful
Guardrail and refusal eventsAttacks, policy issues, overly strict filters

Why Agents Need It More

A single model call is easy to inspect: one input, one output. An agent can make a dozen model calls and tool calls for one request, choose a different path each time, and fail several steps before the visible error. A wrong final answer often starts with a search that returned the wrong document or a tool that returned an error the model ignored. Only a step-by-step trace shows where the run went wrong, which is why tracing is the first thing to set up for any agent, before tuning prompts.

Observability vs Evaluation

LLM evaluation measures quality against a fixed set of cases, usually before release. Observability records what happens with real traffic after release. They feed each other: failed production traces become new evaluation cases, and evaluation scores are attached to production traces to track quality over time.

Common Failure Modes

  • Logging only the final answer. Without the retrieved documents and tool calls, a wrong answer cannot be diagnosed.
  • Personal data in traces. Prompts often contain names, emails and account details. Traces need redaction, access control and a retention period that matches data protection rules such as GDPR.
  • Traces nobody reads. Recording is cheap; reviewing is the work. Without a sampling routine, problems surface only through complaints.
  • No link to versions. A trace that does not record the prompt and model version cannot be tied to the change that caused a regression.

In Practice

Tracing is switched on with the first prototype in our LLMOps services and AI agent development work, because early traces show where a design is weak while it is still cheap to change. LangSmith is the default trace store, with OpenTelemetry export where the client already runs an observability stack, and names, emails and account numbers are masked before a trace leaves the application. A weekly review of sampled traces, with the conversations users rated poorly at the top of the pile, turns each failure found into a regression test.

Back to Glossary

Let's talk about your project

We'll help you harness AI and ML technologies to create innovative software solutions

File requirements: pdf, doc, docx, rtf, ppt, pptx — max 15 MB

No file selected