Glossary
Observability
Being able to see what an AI system did and why: traces of every model call, tool call and result, logs, metrics such as cost and latency, and user feedback, usually collected with OpenTelemetry.
For security it is how a hijacked agent gets noticed and explained afterwards: record the tool calls with their arguments and the outbound requests, and alert on new destinations. Traces also hold prompts and answers, which often contain personal data and secrets, so decide what they keep and who can read them.