An audit trail is the governance ledger: which actor changed which object, what the action was, when, and under whose authority. The system of record keeps the data; the trail keeps the receipts. Compliance reads the trail. Engineers almost never do, until something breaks.
Distinct from an agent-trace. The trace is the reasoning log: what the model saw, thought, and called during a run. You debug the trace. You audit the trail. One explains the thinking; the other proves the outcome to people who weren’t in the room.
Once an agent has write-access, the trail is what makes that write defensible. A person who changes a record leaves a name attached. An agent has to leave at least as much: the action, the approval that authorized it, and the result, all in a form a human can reconstruct six months later without reopening the model. “The agent decided” is not an acceptable answer in a regulated system. The trail is the acceptable answer.
Least glamorous piece of the governance cluster. First thing a security team asks for.