The first generation of AI coding metrics mostly answered whether developers had switched the thing on. That was never going to be enough once agents started consuming meaningful budgets and producing whole changes. At that point usage reporting becomes cost accounting, whether the dashboard calls it that or not.
GitHub has expanded Copilot app usage across its metrics API. Reports can now attribute app activity to individual users and break it down by sessions, requests, prompts, tokens, models, languages, generated code, accepted code, and lines changed. App activity also counts towards daily active users instead of living in a separate total.
That gives engineering leaders a much better view of where the spend is going. It also creates a temptation to turn prompts, tokens, or generated lines into productivity targets. That would be a mistake. A developer who uses one agent session to remove a broken abstraction may deliver more value than someone who generates twenty thousand lines. Activity is useful for allocation and adoption; it is not the outcome.
I would use these numbers to spot expensive workflows, compare tools, and find teams that need help. Then I would pair them with boring delivery measures: review time, defects, lead time, and work that actually reached users. AI costs need visibility, but nobody should get a gold star for burning the most tokens.