Observability

Observability is the capacity to understand a system’s relevant internal state and behavior from designed records, signals, and traces.

Plan for this page

Understand one concept in plain language and connect it to nearby ideas.

You will leave with: For an AI workflow, useful signals can include inputs, versions, retrieval, tool calls, permissions, approvals, outputs, errors, latency, cost, and outcomes. Logging must be proportionate and protect sensitive data.
Time: 3–8 min
Before you begin: No prerequisite. Keep one real example in mind.
Do this now: Read the explanation, then compare the example with the common misunderstanding.

Prepared next stepSee records and monitoring in agent design

ObservabilityObservability is the capacity to understand a system’s relevant internal state and behavior from designed records, signals, and traces.

Recovery
Transparency
AI incident

Plain explanation

For an AI workflow, useful signals can include inputs, versions, retrieval, tool calls, permissions, approvals, outputs, errors, latency, cost, and outcomes. Logging must be proportionate and protect sensitive data.

Example

An operator can trace which model version, document passage, tool call, approver, and external result produced a failed transaction.

Why it matters

Without evidence of what happened, teams cannot diagnose, contest, recover, or improve reliably. Too much unstructured logging can also hide signals and create privacy or security exposure.

Common misunderstanding

Misconception: observability means reading every hidden thought of a model. Correction: it means designing useful system-level evidence around events and outcomes.

Related concepts

Follow relationships, not only the alphabet

RecoveryRecovery is the ability to stop, contain, reverse, retry, compensate for, or safely continue after failure or interruption.TransparencyTransparency is making relevant information about a system, its use, evidence, limits, and governance available to the people who need it.AI incidentAn AI incident is an event where an AI system causes, contributes to, or creates a credible risk of harm, failure, or rights impact.

See records and monitoring in agent design

Sources and scope

What supports this object

Artificial Intelligence Risk Management Framework (AI RMF 1.0)U.S. National Institute of Standards and Technology

Supports measurement, monitoring, documentation and accountability; what can and should be logged depends on system purpose, privacy and risk.

Open source

Loading page content...

…