Agents
Audit Trail
The complete trace of what a system has done: which input, which context, which tool, which output, when, and on whose behalf. Without this trace, you can neither reconstruct an error nor assign responsibility; with it, an opaque workflow becomes a checkable one. For agents in operation the audit trail is not an extra but a condition of approval.
An example
A customer complains about a rejected credit. Without a log, all you can say is that the system rejected it. With a log it says: which model version ran, which instruction applied, which documents were in the context, which tools were called in what order, what they returned, and which rule the rejection hinged on. The difference is that between an apology and an explanation.
A closer look
What is logged is what went in, which tools were called with which values, what came back, and which model version was involved. Without these four items, an incident cannot be reconstructed after the fact.
Why it matters
With non-deterministic systems, logs are not compliance accessories but the only way to find errors at all. You can't reliably recreate a case by repeating it, so what happened the first time has to be recorded. The scope matters: a log containing only input and output doesn't help. What is needed are the intermediate steps, the model version, and the context. And because personal data flows into it, retention period and access rights belong to be settled on the same day as the log itself.
Commonly misunderstood
A log that nobody reads is storage consumption. It only unfolds its value with a question regularly put to it, not after the damage is done.
Where this term does its work in the magazine.
1 article uses the term or takes up its principle.