Logging records what your system did and when.
For LLM systems you log prompts, responses, tokens, latency and cost.
Structured logs as JSON are queryable in a way plain text is not.
Trace IDs let you follow one request across several services.
Never log secrets or personal data without redaction.
Logging is the record of what your system did and when.
It is the first thing you reach for when something goes wrong.
For a language model system, log the prompt, the response, the token counts, latency and cost.
Log the model version and any prompt version too.
Without those you cannot tell whether a change caused a complaint.
Structured logs as JSON are queryable, and plain text is not.
Searching a text log for all slow requests from one customer is painful.
A trace identifier lets you follow a single request across services.
In an agent, log every step: the tool chosen, the arguments, the result.
Agents fail in the middle, and without a trace debugging is guesswork.
Never log secrets, keys or personal data without redaction.
Prompts frequently contain personal data by their nature, which makes this a real obligation.
Redact at the point of logging, not later, since logs get copied.
Set retention deliberately, because keeping everything forever is a liability.
Sample high-volume logs rather than dropping the field you will want.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.