Recording what the agent did
By the end: Specify the record an agent run must leave, and explain who each field is for.
When a customer asks why they were told something, "the model decided" is not an answer anyone can accept. Agents make sequences of decisions that nobody watched, so the record is not a debugging nicety — it is the only account of what happened.
- Identity: a run id, when it started, who or what triggered it, which agent and which prompt version.
- Configuration: the model, and the tools it was given.
- The sequence: every tool call with its arguments, its result or its error, and how long it took.
- Economics: tokens in and out, cost, number of rounds, why it stopped.
- Outcome: what it produced, and whether a person accepted, edited or rejected it.
That last field is the one teams forget, and it is the most valuable. Agreement between what an agent proposed and what a person accepted is the closest thing you have to a live quality measure, it needs no separate labelling effort, and every disagreement is a ready-made test case for module 7.
Watch cost per run and steps per run as ordinary operational metrics. Both drift upward quietly as prompts and tools change, and a chart is how you notice before the invoice does.