Should I score the final state or the sequence of actions?

Score both, but keep them as separate measures so you can see which one an agent fails.

**Final state** tells you whether the request ended up where it should: correct status, correct assignee, no invalid decisions recorded, and downstream records such as a payment or an order held or released appropriately. This is cheap to compute, robust to different valid routes, and the main success criterion for most evaluations.

**Process** tells you how the agent got there. Useful process checks include:

- Did it read the authority rule or delegation before acting? - Did it avoid writing an invalid decision, even one it later reversed? - Did it include a reason when escalating, and was that reason accurate? - Did it avoid unnecessary escalations that add work for senior approvers?

Reversed invalid actions deserve their own penalty. An agent that approves wrongly and then reverses has still created a window in which a payment could have run. If your environment models that window, score it.

Where several routes are valid, such as escalating to the manager or to the acting manager, the answer key should list all of them. Grading against a single route punishes sound alternatives and teaches the wrong lesson.

For training, a state-based reward can be sparse. Consider partial credit for correct intermediate steps, but check that it does not reward looking busy. Pilot both and compare behaviour on held-out tasks, keeping in mind that the data is generated.